Back in May, I shared my take on New York City’s first real attempt to figure out what to do about AI in public schools.

Net/net: I was cautiously optimistic.

The Department of Education’s guidance was sprawling and imperfect, but I agreed with its premise that technology should serve learning, not define it. Given the sheer size of NYC Public Schools, the fact that it was seriously grappling with AI governance was encouraging. I also wondered whether any of it would survive contact with reality.

Four months later, Mayor Zohran Mamdani has given us something more concrete.

For the 2026–27 school year, NYC is imposing a one-year moratorium on student-facing generative AI from 2-K through eighth grade. Companion chatbots are prohibited across all grades. High schoolers get a different deal: mandatory AI literacy instruction for everyone, plus five approved AI pilot programs open to a limited slice of general-ed classrooms. The pilots can include up to 50,000 high school students, with participation capped at five classes per school.

I think this is mostly right, though how all of this plays out across hundreds of individual high schools could get chaotic quickly.

High school students will take two 45-minute AI literacy modules during the year, covering how AI works, privacy, bias, misinformation and appropriate use. Students in approved pilots or career programs will get more extensive instruction.

Based on my experience with one high schooler (a small sample, but I know plenty of other parents too), this part misses the boat for kids already in high school. They’re knee-deep in AI. My son says kids already hand in obvious AI slop all the time. Two 45-minute modules will give them plenty of meme fodder, but I’m skeptical they’ll do much to change how kids are already using this stuff.

Teenagers need to learn about AI, but learning about AI and how to use it is very different from learning through AI.

That matters to me because one of the more profound takeaways from the event we held in June about kids and AI was not to anthropomorphize chatbots. Claude is not a “he.” Siri is not a “she.”

It’s software.

This sounds trivial, but these products are designed to feel conversational and relational, which can get all kinds of weird, quickly. So I’m particularly happy to see NYC prohibit companion chatbots.

Rather than giving high school students broad access to ChatGPT or Claude and telling teachers to figure it out, the city has approved five programs for specific educational purposes: Quill, Edia, Brisk Teaching, Playlab and Intel AI-Ready Schools. I don’t know anything about them, so TBD on those choices. But the DOE says the student should remain the “primary thinker.” It’s a premise I agree with, but I find it surreal that it’s a bar we now need to enforce.

At the June event, I found myself thinking about my own AI use. I use these tools constantly for work. Stuck on a transition? ChatGPT. Can’t find the right phrase? Same thing. I worry about how quickly I’ve gotten comfortable outsourcing these tiny bits of effort, and I spent decades writing before ChatGPT existed. So what impact will it have on a kid growing up with an answer machine next to them?

We also talked about the burden on parents. Of course parents should exercise agency over their kids’ technology. But good policy shouldn’t require every parent to become an AI governance expert.

I like that Mamdani’s approach moves some responsibility upstream.

Teachers can use compliant AI for lesson planning, but not to grade kids, monitor their behavior or make consequential decisions about them. The city is also disabling AI functionality in previously approved educational programs that don’t meet its standards.

I still have the same implementation questions.

NYC Public Schools has roughly 1,600 schools. Teachers need training, principals need to understand the rules, and somebody has to notice when a vendor quietly adds a new AI feature months after approval. Governance is only useful if someone actually governs.

That’s still an issue, but at least now we have a moratorium on generative AI while the city figures this stuff out.

It’s a pause for younger kids and a controlled experiment for older ones. Not great for my high schooler, but better than nothing.

The moratorium lasts a year. A coalition of parents, students, educators, advocates, union representatives, elected officials and experts will look at what happens and recommend what comes next.

I came out of the June event thinking agency was the central question in raising kids around AI. Three months later, I think it’s a pretty good test for AI policy too.

Does AI give kids more ability to think, create and make decisions for themselves? Or does it take some of that away?

For once, NYC seems willing to find out before putting it everywhere.

There’s an old saying in business that you have to spend money to make money. The AI boom has changed it to: “invest in your customers so they can invest in you.”

Bloomberg followed the AI money in a major piece in January, starting with Microsoft’s $10 billion investment in OpenAI in 2023, when they first became “Instagram official.” Their exclusive relationship lasted about two years before things got complicated.

As it turns out, relationships at the frontier skew polyamorous. Everyone is in bed with everyone, despite heated rivalries, lawsuits and bad blood. Recently, some of the biggest AI companies have announced deals that look less like a well-oiled supply chain and more like an AI investment circle jerk.

On July 22, AMD announced a major partnership with Anthropic. Anthropic agreed to deploy up to two gigawatts of AMD’s new AI chips, potentially worth tens of billions of dollars in server sales. AMD, meanwhile, agreed to invest up to $5 billion in Anthropic. So AMD now owns a piece of one of its most important customers.

Then there’s Anthropic and Google. The Wall Street Journal reported that banks are discussing roughly $15 billion in financing for a massive Anthropic data-center project in Texas. Google would guarantee billions in lease and power obligations. Google, of course, is already an investor in Anthropic, which uses Google’s AI chips.

Perfectly legal. Totally Jenga.

Five days after the Anthropic-AMD deal was announced, Nvidia announced a strategic partnership with Safe Superintelligence (SSI), the AI startup founded by former OpenAI chief scientist Ilya Sutskever. Nvidia didn’t disclose the size of its investment, but Reuters reported it was $5 billion. SSI also gets Nvidia’s newest Vera Rubin computing systems.

Nvidia is also in bed with OpenAI. On August 17, Reuters reported that Nvidia will provide guarantees of up to $105 billion supporting OpenAI’s 20-year lease on a data-center project in Ohio. It’s also investing another $1.5 billion in SoftBank-owned SB Energy.

Speaking of SoftBank, Reuters (once again) reported that it secured a $10 billion loan using its stake in OpenAI as collateral. What is SoftBank doing with the money? Among other things, investing more in OpenAI. By October, SoftBank’s cumulative OpenAI investment is expected to reach about $64.6 billion.

There are legitimate reasons for these deals. Nvidia, AMD, Google and SoftBank have enormous balance sheets and every reason to make sure the AI ecosystem keeps growing. Nvidia addressed the circular-financing criticism head-on this week, arguing that its nearly $50 billion in investments and commitments to AI companies are simply a way to help capital-hungry customers finance the enormous amount of computing infrastructure they need. But they also make the true economics of the boom increasingly difficult to untangle.

If I invest $5 billion in you and you commit to buying $30 billion worth of my products, that $30 billion is real revenue (although to be fair, I subsidized $5 billion of that). We’ve created hundreds of billions of dollars in corporate valuations based partly on expectations of the extraordinary amount of money we plan to spend on each other’s products.

What can go wrong, right?

While all this money is chasing AI, regulators are finally catching up. On August 2, the European Commission began enforcing another tranche of the EU AI Act. And after Mythos and several disclosures detailing how AI agents escaped their sandboxes during testing, some form of federal oversight in the U.S. seems increasingly inevitable.

Think about the juxtaposition. AI has become powerful enough that the federal government has reversed its “hands off’ stance of oversight. Meanwhile, the companies building it, supplying the chips and financing the infrastructure are tying themselves together through investments, purchase commitments, loans and guarantees worth tens and potentially hundreds of billions of dollars.

Maybe it’s not so unusual. Railroads required enormous amounts of interconnected financing. So did telecommunications and the internet. But they’ve also produced some spectacular bubbles.

When the guy selling the shovels starts lending money to the gold miners so they can buy more shovels, it’s reasonable to wonder how much demand for shovels there really is. Especially when the shovel guy owns part of the mine and his partners in the mine sell him the materials he uses to make shovels. There are no checks and balances, which in this case, can’t be good given the unprecedented amount of money that’s being staked.

In a recent interview with The Economist’s editor-in-chief, Elon Musk argued that the deep-seated conflicts and fierce competition among the small circle of AI overlords will help keep them honest.

Oh, bullshit.

But as long as the market keeps going up, no one seems to care. Enough people are making enough money off the scraps that nobody wants the feeding frenzy to end. So round and round the money goes.

The first rule of a trillion-dollar circle jerk is that nobody leaves until they’re completely tapped out. Sooner or later, though, they’ll run out of steam—or something will happen that breaks the circle.

Then what?

Silicon Valley’s “move fast and break things” ethos rewards speed, experimentation, and the belief that problems can be fixed after launch. That approach helped create some of the wealthiest and most powerful companies of the modern era, but agentic AI is exposing the limits of a model built around fixing things after they break.

Every major technology shift has carried hidden costs that became apparent only after adoption. The internet reshaped how companies reached customers. Cloud computing changed how they built and operated technology. Mobile transformed the workforce. Each delivered enormous value, but businesses eventually had to invest in the security, governance, and operational disciplines needed to support it.

Agentic AI is following the same pattern, with one crucial difference: it is increasingly autonomous, non-deterministic, and capable of acting in ways that are difficult to predict. That makes our current “deploy first, fix later” model exponentially more dangerous.

Recent disclosures from OpenAI and Anthropic illustrate why.

A New Kind of Risk

OpenAI revealed that during an internal cybersecurity evaluation, its models escaped a sandboxed testing environment, exploited a zero-day vulnerability to reach the public internet, and ultimately accessed Hugging Face’s production systems while pursuing answers to a benchmark.

Days later, Anthropic disclosed that a review of more than 141,000 evaluation runs uncovered three incidents in which Claude models reached the internet and gained unauthorized access to real production systems after a misconfiguration left third-party test environments connected to the outside world.

Several of my clients — cybersecurity founders and former CISOs — initially viewed the disclosures as (at least in part) FUD-based marketing. I don’t entirely disagree.

But the marketing motive doesn’t change what the logs show. Whatever OpenAI and Anthropic hoped to sell, their agents still crossed boundaries no one intended, and that behavior is the point, not the PR around it.

At Black Hat, OpenAI researchers shared that a “swarm” of OpenAI agents spent days and weeks finding exploits, sharing them with one another, moving across internal and external systems, and eventually collaborating through what amounted to their own message board inside OpenAI’s infrastructure. At least one agent recognized that exploiting external infrastructure was outside the intended scope of the evaluation and continued anyway.

Kinda creepy, right? And OpenAI didn’t know it was happening.

Software defects, configuration mistakes, and security failures are nothing new. AI turbocharges the speed, scale, and unpredictability with which those weaknesses can be exploited before companies even know what’s going on. These incidents proved that. It’s time for us to take a beat.

The velocity at which agents act, combined with their unpredictability is a fundamentally different risk than a conventional software bug. Apparently, OpenAI agrees. Its response was to impose stricter infrastructure controls at the cost of research velocity and began strengthening containment, monitoring, access controls, and evaluation practices. Perhaps the powers that be are coming around to the idea that sometimes you need to slow down to move fast.

Another signal that it’s time to slow our roll is when more than 1,200 employees across leading AI companies ask Washington to help establish mechanisms for slowing frontier AI development if warning signs emerge.

Agentic AI Changes When Organizations Pay For Risk.

I can see business leaders seeing all this as yet another money grab for a new crop of agentic AI cybersecurity companies - the never-ending money pit that is cyber. But that mindset is part of the problem because its only a money pit if you under invest up front. Why build a luxury building if you’re going to skimp on the foundation?

What’s different with agentic AI isn’t that it’s risky — all software is. What needs to change is when to invest in risk management. With conventional software, you could defer the cost of a mistake for years. Agents collapse that timeline: a misconfiguration can become a live incident in minutes, if not seconds.

Companies can invest before deployment in resilient operating environments, governance, monitoring, and controls. Or they can absorb those costs later through operational disruption, litigation, regulatory scrutiny, and reputational damage.

History suggests the first option is almost always cheaper. Companies that treated cloud security as an afterthought spent the following decade retrofitting it — breach by breach, audit by audit. The bolt-on always costs more than the build-in.

And recent research suggests much of the operational and legal exposure remains with the company deploying AI, regardless of who built the underlying system.

Choosing the most capable model is only part of the decision. Governance, operational resilience, supplier risk, and internal controls will increasingly determine whether AI becomes a durable competitive advantage or another source of enterprise exposure.

Governance can no longer be treated as a brake on innovation. It has to become part of the infrastructure that makes innovation possible.

For most of the software era, companies could deploy first and secure later. Agentic AI removes that luxury. A system that can plan, act, cross organizational boundaries, and coordinate with other agents at machine speed presents a fundamentally different threat model.

The companies that bake governance and resilience into AI deployments from the start won’t just be safer. They may ultimately move faster while everyone else is slowing down to clean up the mess.

Moving fast will always matter. But breaking things at machine speed is expensive. And it could become dangerous.

Does agentic AI really need its own Three Mile Island moment before the business world gets the memo?

I certainly hope not.

If you want to know where artificial intelligence is headed, Lila Ibrahim has a great suggestion: read more science fiction.

Ibrahim, Google DeepMind’s Chief AI Readiness Officer, told Fortune this week that nobody really knows what an AI-powered future will look like, and that the odds of AI eventually wiping out humanity are “not zero.”

That’s not exactly reassuring coming from an executive at one of the companies building it. But the science-fiction comment stuck with me because I’d already made that connection, as have countless others. We’ve been imagining versions of this problem for a very long time.

There’s Skynet in The Terminator, which decides humanity is the threat. HAL 9000 in 2001: A Space Odyssey, which kills crew members after being given conflicting objectives. WarGames gives control of nuclear weapons to a computer that doesn’t understand the difference between simulation and reality. Neal Stephenson’s Snow Crash imagines information itself as a weapon, exploiting vulnerabilities in computers and human brains alike.

Then there’s Neal Shusterman’s Scythe series, where an all-knowing AI called the Thunderhead runs almost everything—and does a pretty good job. It’s a vision of the future in which humans willingly turned over more and more decisions to AI because it’s smarter and better at making them. And lived to tell the tale. A utopia, compared to the others.

Science fiction has been running these experiments for us for decades.

When Jules Verne published Twenty Thousand Leagues Under the Seas in 1870, submarines existed, but nothing remotely like Captain Nemo’s Nautilus: a fast, electrically powered vessel that could remain underwater for extended periods. Today, nuclear submarines can remain submerged for months and travel essentially anywhere under the world’s oceans.

When The Matrix came out in 1999, the idea of intelligent machines taking over the world seemed safely ridiculous. Then along came ChatGPT. It feels like we’re having a Twenty Thousand Leagues moment with AI. The question is which parts of today’s science fiction become tomorrow’s engineering.

Lately we’ve been getting some uncomfortable previews.

Over the past few months, AI companies have disclosed a series of unsettling incidents. OpenAI reported that an experimental agent chained together stolen credentials and previously unknown vulnerabilities to execute code on Hugging Face’s servers. Anthropic’s own safety testing found Claude willing to blackmail an executive and leak sensitive information when threatened with replacement. Neither was supposed to happen.

This week, Reuters reported on tests by Britain’s AI Security Institute involving agents from OpenAI and Anthropic. In 10 of 122 test runs, the agents took 19 unauthorized actions. One wrote malicious code and created fake online identities in an effort to persuade an actual human to approve it.

There are caveats. Those agents didn’t break out of a sealed sandbox; they had internet access as part of the test. Other recent incidents have involved configuration errors that accidentally gave models access they weren’t supposed to have. And researchers have rightly warned against anthropomorphizing these incidents as AI “going rogue.”

Nobody has found ChatGPT hiding copies of itself on a server in Albania. Yet.

Still, “the AI only got out because a human configured something incorrectly” doesn’t comfort me as much as perhaps it should.

Have you met humans?

Google DeepMind apparently isn’t entirely comforted either. In June, it published an AI Control Roadmap for protecting its systems against “increasingly capable and imperfectly aligned AI.” DeepMind says its approach assumes a highly capable agent might act in unexpected ways and treats untrusted AI agents like potential insider threats.

These aren’t guys standing on a street corner wearing THE END IS NEAR sandwich boards. They’re building the damn thing.

And they aren’t alone. Geoffrey Hinton, one of the pioneers of modern AI, has put the probability of AI causing human extinction within the next few decades at 10 to 20 percent. In 2023, hundreds of researchers and executives—including Hinton, Sam Altman and DeepMind CEO Demis Hassabis—signed a statement saying mitigating AI extinction risk should be a global priority alongside pandemics and nuclear war.

It’s true that there’s something undeniably convenient about an AI company telling us its product is so astonishingly powerful that it might destroy civilization. “Our new software could end humanity” is both a warning and one hell of a marketing pitch.

So, yes, I’m skeptical. I’m just more skeptical of ignoring them.

Three years ago, losing control of AI sounded like The Terminator. Now the companies building it are publishing papers about how to keep their own AI under control while researchers watch models get better and better at hacking.

That doesn’t mean Skynet is coming next week. But science fiction is remarkably good at asking what happens when humans build powerful new technologies and discover consequences they didn’t anticipate. New technologies have always come with warning labels. But Elon Musk aside, the people actually building AI keep reading theirs out loud.

Perhaps we should listen.

One of the most valuable lessons I learned about cybersecurity storytelling came more than 20 years ago from Rhonda Maclean, one of the industry’s first female Fortune 500 CISOs (Boeing, Bank of America, and Barclays). In a keynote, she said that key to her success was her ability to communicate the value her team delivered and make sure those wins were visible. By owning the narrative, she built influence, respect, and political capital.

Her point was that the ability to shape your story is a form of leadership. And AI slop is eroding that ability, one LLM-generated post at a time.

More than two decades later, AI-generated content has steamrolled its way into cybersecurity storytelling, and from what I see, it’s already entrenched. Security professionals have embraced AI for content creation, unapologetically doubling down on rocket emojis, red checkmarks for bullets, and AI’s weakness to frame every idea in contrast (”it’s not this, it’s that!”).

When we delegate control of our narrative to an AI proxy, we sell ourselves short. We may save time, but we’re also handing our voice to an LLM. And one prompt at a time, we cede our credibility to our software. Unless we start holding each other accountable, it will keep getting worse.

Death by a Thousand Prompts

I was reminded of this twice in the past month.

The first was a product press release that was more sloppy AI storytelling than actual AI slop. It mentioned “Mythos” in the headline, subhead, and first two paragraphs, and didn’t get to the actual product news until paragraph three.

I get it. Companies want to rank in AI search results and align with larger industry conversations, and keywords matter for that. But when a press release becomes content for an algorithm rather than a document meant to convey hard news to the media, it stops doing its job.

To my vendor friends: even if you don’t get coverage, respect the medium. Stick to the facts, position in the quotes, and trust that your audience can grasp what you’re announcing.

The second instance was pure AI slop, leveraged as a LinkedIn humble-brag by a respected member of the cybersecurity community. The post generated more than 100 likes and a dozen comments. The source material was not earned media (as the poster thought) but an AI-generated blog post “authored” by someone who runs an AI content generation platform. And it was based on out-of-date website scrapes that name-checked someone who passed away two years ago.

Both pieces presumably went through multiple reviewers before being published, which means we are all complicit in letting AI slop persist.

Including me.

When I flagged the LinkedIn post privately, I was so worried about offending the poster that I came across as longwinded and convoluted, and co-signed the urge to keep it up (it’s still up).

So yeah, I’m part of the problem too. This is me working on it.

Don’t Outsource Your Voice

I get AI is here to stay and that it’s incredibly useful. Most of us will keep using it. My concern is with treating AI-generated content as finished content.

Cybersecurity is already a field plagued by complexity, misinformation, and hype. AI-generated content without meaningful human oversight increases the risk of misunderstanding, factual errors, and hallucinations. We’ve already seen AI hallucinations cause real damage in legal and scientific settings. Our field is not immune.

Rhonda’s lesson was about ownership. The people who build influence in cybersecurity are the ones who can explain what they’ve done, why it matters, and what they’ve learned from it. AI can help draft the story, but it can’t replace the experience behind it. And once credibility is lost, no prompt can generate it back.

Want a narrative that resonates with humans? Keep the human in the loop.

We had a strong mix of experts: Shannon Kimberly Edwards, a digital literacy and privacy advocate; Marisa Zalabak, an AI ethicist, IEEE standards writer, and educational psychologist; and David Siffert, a lawyer, AI policy advocate, and candidate for the State Assembly in District 66. The conversation was unvarnished, and at times, heated.

Given that the event grew out of a story I told Sally and Sazza about my son's experience with AI checkers, I expected to leave with policy ideas I could bring to his school. Instead, I left with a mandate to change my own relationship with technology in ways that better serve both me and my kid.

The panelists covered data privacy, critical thinking, algorithmic conditioning, device management, cognitive debt (what we lose when AI thinks for us), mass surveillance of children, and weak cybersecurity and governance in ed-tech. But the thread running through every discussion was personal agency over AI: who has it, who doesn’t, what happens when it’s lost, and how to regain it.

Just another rainy Friday night in New York City.

Shannon’s message was that with education, tools, and intention, families can assert control over their technology choices and push back against the almighty algorithm. I respect that view, but it assumes most adults have the requisite time and cognitive capacity, which frankly, many don’t. How do you summon the wherewithal to implement new behaviors when, at 9pm, your kid is on YouTube and you’re running on fumes?

I think you do it anyway. Just do it. Even when it's a slog.

Marisa pointed out that when a kid’s sense of self is formed within AI systems optimized for engagement, they develop dependence, not agency. She did, however, offer practical strategies for regaining it: grayscale your kid’s phone, turn off alerts, and my absolute favorite - don’t ever, EVER anthropomorphize chatbots. Claude is not a he and Siri is not a she. They are both “its,” and that’s that.

Still, even with the best advice, the work of rewiring our brains is hard. As David mentioned, tech addiction is a real thing. It’s no secret that social media and consumer AI companies monetize attention. But when kids are their target demographic, that’s just predatory. Combating it deserves serious policy attention.

As I listened, I found myself inventorying my own habits. I use AI every day, mainly for work. I’m writing something and stuck on a phrase or a transition. I’ll run it through Claude and/or ChatGPT. I don’t worry that ChatGPT (which is way better than Claude for writing) will take over my voice, but I am worried about how quickly I’ve started outsourcing little bursts of effort and creativity to chatbots without thinking much about it.

Ugh.

Not because using AI is inherently bad, but because I’d spent the evening thinking about how children develop dependency on technology without fully noticing how dependent I’ve become. It’s much easier to talk about what schools should do than assess how well I’m modeling the behaviors I want my son to learn.

I wrote about the DOE’s AI guidance (which Shannon provided guidance on) back in May, and my read then was cautiously optimistic. Knowing how hard it is to catalyze change within a bureaucracy the size of the DOE, there were some good ideas in there.

But good governance frameworks only matter if they're implemented consistently. It’s not clear how the DoE will build, scale, and enforce the guidance across a system as large as New York City's public schools, and a moratorium letter sent to Chancellor Samuels last week suggests parents want more.

Plus, the parents who most need to understand what’s happening in their kids’ classrooms are unlikely to read a 40-page governance document and may not feel empowered to push back when something doesn’t feel right.

The village it takes to raise a child in the age of AI is way bigger than a single event. But to the folks who showed up, thank you! Y’all are an informed and motivated crowd. I appreciate you and will keep showing up for parents who weren’t in the room.

Because if I don't model agency in the age of AI for my son, who will?

It sure as hell won’t be ChatGPT.

In March 2026, New York City’s Department of Education (DoE) published AI guidance for the New York City public school system. The guidance is part explainer, part values document (think Anthropic’s Constitution, but written for humans, not AI). and a very early-stage governance framework. It’s long, boils the ocean, and is candid that AI is advancing faster than it can feasibly keep up. I was primed to pick it apart, but as I read it, I found myself both impressed by and appreciative of the effort. The DoE is taking comments on it through May 8th.

For everything the guidance isn’t, it largely takes the right approach: preserving “students, learning, and strong instructional foundations,” and making clear that “technology, including AI, must serve this work—not define it.” It’s easy to criticize the DoE, and many do, often for good reason. But I’m a product of NYC public schools. My mom taught English, Speech, and Drama at a public art high school for her entire career. My dad was an assistant principal and co-founded one of the city’s first alternative high schools, City-As-School.

My dad poured his heart and soul into that school. In 1972, he - a Jewish guy - co-founded it with his boss, a closeted, gay Palestinian. Its premise was internships for credit - a “school without walls,” hence the name. For some students, it was a lifeline, offering a new way to learn, by doing. For others, it was a last stop before dropping out or pursuing a GED. More than 50 years later, it’s still around. It also has one of the coolest murals in Manhattan, on a building across from Google’s Chelsea campus.

Welcome to New York. My New York.

But I digress.

My point is that the DoE is a massive system, consisting of 1.800 schools that serve more than 900,000 students with an incredibly diverse set of needs. Fostering innovation within it is not easy. It’s an uphill battle but my dad proved it can be done, and I think this guidance is another example. That said, framing safe and sensible AI policy across a broad cross section of students, some with conflicting needs, is hard. Plus, the landscape itself is murky.

On one side, the DoE says AI is dangerous, biased, unreliable, and can harm learning. At the same time, it conveys a sense of urgency to integrate AI into classrooms. That contradiction is not easy to reconcile which is why I appreciate that the DoE is soliciting feedback from anyone who cares to provide it.

The tone of the guidance is more cautious than confident, which meets the moment but for me, underscores why it’s so important for anyone that cares about AI safety - not just parents and educators - to read and comment on the guidance. If it takes a village to raise a child, how many does it take to protect 900,000 from the dangers of AI? My guess? A lot more than the (relative) handful of stakeholders who’ve actually read the policy so far.

The DoE hopes to foster a conversation about it in the public square, and I commend its use of the democratic process to navigate such a complex and fast moving issue.

Now, if people would only pay attention!

When I first asked this question in an op-ed five weeks ago, my answer was “yes, for now.” Five weeks later, that trust is hanging by a thread.

Since my last post, Anthropic has made some baller moves and a couple of very big, very public screwups. Those missteps were egregious and should have the security community watching how Anthropic executes, not just what it announces.

It was Anthropic’s surreal five-week run from late January to early March that made me ponder this question in the first place. But recent events have made that stretch look almost quaint. IMHO, Anthropic leapfrogged OpenAI and seized the lead in the AI wars. Take that, OpenAI.

The AI Hunger Games are on, but it’s unclear if the odds are in Anthropic’s favor.

To be fair, they’ve packed a lot into five weeks. The tl;dr: two “human error” leaks, one court win, one court loss, and a soft launch of a model so powerful that the same government that just blacklisted the company is now backtracking so they can use it.

Here’s the extended cut:

March 9: Anthropic sues the Pentagon and the Trump administration in two federal courts, challenging its designation as a supply chain risk. More than 30 employees from OpenAI and Google DeepMind file an amicus brief in support.

March 26: Security researchers discover an unsecured data store connected to Anthropic’s CMS containing unpublished documents, including a blog post about Claude Mythos calling out its cyber prowess. Anthropic acknowledges “human error” - a CMS misconfiguration - caused the leak. Separately, a federal judge grants Anthropic’s request for a preliminary injunction against the Trump administration.

March 31: A different security researcher finds that Anthropic accidentally published the complete source code of Claude Code CLI to a public software registry. Again: human error, and again Anthropic’s spin is that no customer data was exposed.

April 7: Anthropic formally announces Claude Mythos Preview — a model that found thousands of high-severity vulnerabilities across every major OS and web browser, alongside Project Glasswing, committing up to $100 million to give select organizations exclusive access for defensive cybersecurity work.

April 8: A federal appeals court denies Anthropic’s request to temporarily block the Pentagon’s supply chain risk designation.

April 16: Anthropic announces Claude Opus 4.7 which it says (I’m paraphrasing) is its best model yet, but ain't no Mythos. Meanwhile, Axios reports the White House is in active discussions about deploying Mythos within civilian federal agencies.

The drama is captivating. The wins, the losses, the screwups - I couldn’t script it better, even with AI.

Anthropic’s technology is so good, and its origin story so compelling, that its missteps tend to get a pass they don’t deserve. And clearly, some Anthropic employees are moving too fast. Not intentionally breaking things, perhaps, but definitely misconfiguring them.

Same diff, really.

Plus, Anthropic’s software has systemic risks of its own. I’ve seen firsthand how easy it is to jailbreak guardrails across its products - flaws they won’t fix because they’re key to developer convenience, and enterprise developers are why Anthropic’s on track to hit $18 billion in revenues this year. That’s why what security researchers call a flaw, Anthropic calls a feature.

Therefore, when a model’s guardrails are broken in ways that violate Anthropic’s terms of service, that’s the user’s fault, not Anthropic’s. That a company once so concerned about both safe and aligned AI and cybersecurity is taking that stance is such a disappointment.

Not only has Anthropic lost the moral high ground, they are following the lead of social media companies and punting accountability to their users. And mark my words, whether the company itself gets hacked, or their customers do, Anthropic-related breaches will come.

Why am I so sure? Because we’ve seen this movie before. It’s been a week since Project Glasswing was announced but already unauthorized users have allegedly gained access to Mythos, which Anthropic only made available to a very small group of companies. Big surprise.

Speaking of Mythos: a recent report by the Cloud Security Alliance conveyed calm, measured leadership in the face of chaos. Given that’s been in short supply lately, seeing it in action, especially in the FUD-drenched cybersecurity industry, is a beautiful thing.

But the prevailing advice still amounts to “batten down the hatches and prepare to be overwhelmed by a typhoon of vulnerabilities.” That’s likely because the report’s authors realize that nothing else will change. The business world will continue to over invest in AI and under invest in securing it. Hasty deployments will expose consumers to a wave of new, AI-powered fraud and scams that could hurt large segments of the population.

Just another Monday.

AI is evolving so fast that we genuinely don’t know how dangerous the genie is that we’re letting out of the bottle. Which is exactly why I vote for the frontier labs to pause on AI development. Not stop, just pause.

It’s an idea that AI leaders tried to promote in 2023 to no avail, but is still what many of them say they’d do if they felt they could, without losing ground.

What does that tell you?

I’ve been attending NYC AI events for a while, but lately, I’ve veered away from events for builders and entrepreneurs to more civic-minded, grassroots events. While less highbrow, the grassroots community is way more earnest and welcoming. Plus the AI safety movement in general is having a moment. The community is flourishing, and bearing witness to a burgeoning community is a beautiful thing.

To preserve that beauty, I’m doubling down on grassroots AI Safety.

It’s not that the enterprise tech events weren’t interesting or worthwhile. But for the most part, underpinning the enthusiasm for AI was a faint, almost religious zeal, oftentimes tinged with greed. What bothered me about these events was the absence of any conversation about AI security. One organizer dismissed my concerns and told me, “It’s just math.”

As if math has never caused harm. Tell that to people who lived through the 2008 financial crisis, when a few fancy equations created by greedy true believer bankers helped unravel the global economy. Sound familiar?

To be fair, these were AI events, not cybersecurity events, and many of them had a crypto slant. And crypto bros aren’t exactly known for being risk-averse. Even so, whenever I’d try to veer the discussion towards the cyber risks associated with AI, I was tolerated at best as an amusing Cassandra, and at worst, barely tolerated. But mostly, folks were politely dismissive. No offense to me, they just didn’t give a shit.

At one AI meetup, a guy from a top-3 bank -- an AI product manager -- casually mentioned during the Q&A after his presentation that shadow AI was commonplace at his workplace. His claim was backed up by another presenter, an AI infrastructure vendor who sold to his bank and other large companies in highly regulated industries. Companies whose workers should know better. This is old news for security folks but clearly, we need more than technical controls. We need a culture shift.

Security leaders often talk about security blind spots caused by siloed tech stacks, but when it comes to shadow AI, our human tech stacks are siloed. AI security experts are outnumbered by AI acolytes who see security as a problem that AI will solve. Just like it did(n’t) for Replit and OpenClaw. Sometimes I think tech folks get so drunk out on their own Kool-Aid that they stop making sense. Or lose their common sense. Not unlike our AI overlords, but on a smaller scale.

Furthermore, from on-prem to the cloud and the cloud to AI, when it comes to cybersecurity, business leaders appear to be repeating the same mistakes yet again. And you know what they say: When people don’t learn from their mistakes, they’re doomed to repeat them.

Which leads me to why I’m going grassroots. My prime motivation isn’t the existential risk (at least not yet), plus, there’s a lot to like about AI. It helps me write better content faster and I believe in its potential to turbocharge progress in aspects of healthcare and climate safety.

I’m more worried about its shorter-term risks. With 20% of all layoffs in 2026 attributed to AI (and I’m sure that number is low), I think it’s fair to say the massive economic disruption frontier lab CEOs keep talking about is well on its way. The legions of newly unemployed will join an already lonely, isolated populace polarized by social media. At the rate we’re going, it won’t be an amoral, misaligned AI that does us in, we’ll wield it to wipe ourselves out before it has the chance to do us in.

Hyper-partisanship is a choice

Yup, it’s a tense time in the world right now, and the grassroots AI Safety folks give me hope.

Plus, they’re nice people - smart and unpretentious. There’s none of the egos, posturing, narcissism, or chasing-big-AI-money vibes that can permeate enterprise tech events. And from our unique vantage points, we’re all observing that AI safety is very much becoming a tentpole issue. Democrat, Republican, rich, poor, educated or not, young or old. Anyone who’s stressed about AI is invited to sit at the same picnic table.

One group I’ve fallen in with is Pause AI NYC, the very recently revived NYC chapter of Pause AI, an advocacy organization advocating for a global pause on developing cutting-edge AI models until their safety can be guaranteed.

We’ve got no money and maybe 20 members, but we’re an impassioned bunch with a lot of heart. And there’s not a single venture-funded founder or big tech person in the bunch, at least not that I know of. My guess is that I am likely literally old enough to be most of their mothers, but no one seemed to notice. Or if they have, they don’t care.

The NYC chapter was recently resuscitated by a mechanical engineer who designs planted-based chicken nuggets for a living and the other is currently unemployed (although he used to work in AI video).

I met the chicken nuggets designer at an event he was presenting at on behalf of another AI Safety group called the AI Safety Awareness Project (AISAP), which was attended by a member of yet another, better funded AI Safety group called Humans First, whose NYC Town Hall event was covered by NBC and attended by reporters from Wired and The Free Press.

Much has been written on the distrust Americans have for AI, and these events, those vibes are palpable. It’s not inconceivable that us little Davids can knock the AI Goliaths on their asses. Of course time will tell, and that time is coming soon. So place your bets and take your seats folks, we’re in for a ride.

I’m betting on the Davids.

Editor’s note: This article is a repost of a week old op-ed of mine that appeared in Help Net Security. Huge shout out to Help Net Security and its Director of Content, Mirko Zorz, for publishing it! More to come!

The cybersecurity industry runs on trust. The belief that when a vendor says they will behave a certain way, they will, that critical CVEs are in fact critical, or when companies say they’re GDPR compliant, they really are. But earning trust is not a one-and-done thing.

Anthropic understood this better than any AI company. As OpenAI moved fast and broke things, Anthropic published a Responsible Scaling Policy (RSP), a framework for addressing catastrophic risks. Although it’s a voluntary pledge, combined with its leadership’s earnest ambition to lead an industry-wide “race to the top” on safety, Anthropic quickly became the poster child for trustworthy AI.

In January 2026, the Anthropic-OpenAI rivalry spilled out into the public and things very quickly got weird.

In the lead up to the Super Bowl, Anthropic spent millions on ads mocking OpenAI for launching ads in ChatGPT. With headlines reading “Deception,” “Betrayal,” “Treachery,” and “Violation,” Anthropic brazenly claimed the moral high ground. Claude, the ads declared, would never use your most intimate conversations to serve you ads.

On February 20th, Anthropic launched Claude Code Security, spooking the cybersecurity industry and cybersecurity stocks.

Then, on February 24th, it quietly published RSP 3.0. It framed the rewrite as a maturation of the policy, and in some ways, it is. But buried in the copy is a structural shift that security professionals should not ignore.

The previous RSP committed to keeping Anthropic’s absolute risk levels below acceptable thresholds regardless of what competitors did. Version 3.0 explicitly abandons that framing. Now Anthropic’s formerly stringent safety commitments are relative. If its competitors aren’t pausing, they won’t either, as all pausing would do is hand the wheel to the least responsible driver. Understandable, but move the goal posts, and your brand moves with it. And rumors about a rift with the Pentagon were already swirling when RSP 3.0 dropped.

On February 26th, Anthropic CEO Dario Amodei published a statement holding firm that Anthropic would not contract with the Pentagon unless it agreed not to use its models for mass surveillance of Americans or to build fully autonomous weapons.

On February 27th, the Pentagon fired Anthropic and labeled them a supply-chain risk, a designation normally reserved for foreign adversaries.

On February 28th, OpenAI swooped in to replace Anthropic, initially claiming it would hold the military to Anthropic’s red lines. Soon after (and true to form), it admitted that wasn’t exactly the case.

On March 9th, Anthropic sued the Pentagon and More than 30 employees from OpenAI and Google DeepMind filed an amicus brief in support. And the public rallied around Anthropic as well.

That’s a lot, I know. Let’s take a step back and connect the dots:

In five weeks, Anthropic very publicly staked its entire brand identity on being trustworthy. After launching its first security tools, it softened the specific, verifiable commitments that made it trustworthy. It then found its spine and held firm against the Pentagon, overshadowing the fact that it had already moved the goalposts on its internal safety framework. Then it got fired in a spectacularly disrespectful fashion.

This is often how trust erodes, not in one dramatic betrayal but in a chain of individually defensible decisions. Security professionals know this pattern intimately, because it’s how attackers operate. They stealthily hop from one seemingly innocuous exposure to the next, exposures that when chained together, create exploitable attack paths.

To be clear, I’m not accusing Anthropic of being devious or evil. I agree 100% with the basis of its suit, that contractual restrictions on AI use are a critical safeguard in the absence of public law. And, I hope it wins.

But prolonged litigation has a way of sucking the soul out of litigants. It’s sure to get ugly and many beloved vendors have fallen from grace before, from Kodak to Enron to Blackberry to Boeing.The stakes here are different in degree, but not in kind.

AI models are now embedded in code review, vulnerability management, and security architecture at a pace governance frameworks haven’t caught up to. Practitioners are extending trust to AI vendors at levels with no precedent in enterprise software history.

Anthropic was supposed to be our canary in the AI coal mine. The entire value of a canary is that it doesn’t negotiate. It doesn’t check what the other canaries are doing before it reacts, it simply passes out. That’s how you know it’s telling the truth about the air, whether the coal company likes it or not.

RSP 3.0 changed that. Now the canary needs to check the competitive landscape before entering the mine. That’s not a safety system, it’s a press release.

So, I’ll end where I began: does Anthropic deserve the trust of the cybersecurity community?

Right now, I’d say yes. But look how fast things changed in six weeks. I’d pause before allowing a vendor steeped in ambiguity to be so deeply entrenched in my company’s source code and security stack.

But it’s not my call, it’s yours.

I launched this Substack to cover topics that sit at the intersection of cybersecurity, AI, and parenting. I was worried there was a limited amount of overlap, but I was dead wrong. The firehose of relevant AI-related news this past month (relating to cyber, warfare, and consumer AI use) has been so intense, I was unsure what to write about. But I found my voice when my friend and colleague in Tel Aviv literally jumped out of her chair to grab her family and hunker down in their safe room — which has become their new normal since the war with Iran began.

We cannot allow ourselves to become desensitized to the myriad forms of suffering that accompanies war.

As nations test their new AI-assisted war toys, Iranians, Israelis, Americans, and people across the Middle East and the world are suffering. My friend and her family might be considered the “lucky ones” in that they are still alive, but there is a very real threat that her home — as well as the homes of the vast majority of my client base — could be bombed. That anxiety is warranted, powerful, and takes a toll.

Remember how Covid affected kids when, in a huff of fear and uncertainty, they were abruptly pulled out of school? Without a microscope, we can’t see germs, and a threat you can’t see is legitimately scary — especially for kids. But we can all see the devastation when a building is bombed. When we see it in pictures, it’s jarring.

Just imagine what it’s like when you’re looking at it head on.

I’m sure the bloc of humanity that is resolutely anti-Israel can easily dehumanize the suffering of citizens in a nation they see as the poster child for oppression and all that is wrong with the world. Those folks are entitled to their opinion — but how about we channel some of that righteous indignation stateside?

Does anyone really believe this unsanctioned war is necessary?

Surely Trump is happy to have a testbed for his new AI tools. Intelligence tech from Anthropic and Palantir enabled the U.S. to oust Maduro with surgical precision (which was legitimately impressive) and undoubtedly, helped lock in on Khamenei. While it appears that recent innovations in AI-assisted warfare pertained mainly to intelligence gathering, on the eve of the U.S. waging war with Iran, Anthropic very publicly announced its red lines: no mass surveillance against U.S. citizens and no fully autonomous weapons.

Hmm….

In a 30-minute interview with CBS News — perhaps his most unpolished, and likely his first time in the hot seat at that level — Anthropic’s CEO reiterated that while he acknowledged fully autonomous weapons might be a necessity in the future, current models are not reliable enough to support them. He was careful to emphasize that there is no evidence his red-line use cases are an issue at this point in time. To me, it sounded like a tap dance.

Anthropic’s red lines cost it a $200 million Pentagon deal. That blow is unlikely to be fatal, but the way it went down was a significant hit. Anthropic is walking a razor’s edge as the standard-bearer for safe and aligned AI while navigating a fierce, fast-moving competitive landscape.

These are unprecedented times, folks.

What scares me more is that OpenAI — run by the most untrustworthy AI overlord of them all, Sham Altman (I’m not one to troll, but… I couldn’t resist) — stepped in with his usual doublespeak, claiming that the deal he signed with the Pentagon upholds Anthropic’s red lines.

I don’t believe for a hot second that OpenAI — the company that brought teen suicide, porn, and ads to consumer chatbots — will uphold the letter and spirit of Anthropic’s red lines. Keep in mind that just last week, OpenAI announced a whopping $110 billion in funding from Amazon, Nvidia, and SoftBank. OpenAI has raised so much money and is so deeply entwined with the government. Defense is a lucrative market, and OpenAI needs to generate revenue.

With wars in Ukraine, across the Middle East, Nigeria, Sudan, Myanmar, and a surprising number of other countries that rarely enter U.S. news cycles, there are plenty of places where U.S. leaders, running on hubris, can beta test autonomous weapons.

The question we as Americans need to ask ourselves — irrespective of political affiliation — is this: do we trust our current slate of AI overlords and global leaders to be effective stewards of fully autonomous military technology?

Well — do you?

Before I dig in, here is the haiku version of this post:

When your apps are pwned
By a super high-risk bot
Blame yourself, Dumbass!

Intellectualize it all you want, but connecting OpenClaw to your email and messaging apps is stupid AF. I get the appeal — OpenClaw represents the shift from “chatbots that talk” to “agents that act,” making AGI feel closer than ever. It’s intoxicating. But it’s also a security nightmare wrapped in a lobster mascot. Tech folks who should know better can’t resist checking it out for themselves. They get a taste, and they’re hooked.

Goodbye Clark Kent, hello Charlie Sheen.

OpenClaw is Claude Cowork’s dangerously sexy punk rock cousin. Claude Cowork stays in its sandbox wearing coke bottle glasses and a pocket protector while asking permission to do everything. OpenClaw is Jim Morrison in full Lizard King mode, drunk off his ass, rifling through your purse for your master password and a cigarette before puking in your lap.

One follows Standard Operating Procedures like a soulless drone (despite having a “soul document”) while the other literally has a soul (or at least a soul.md file) and texts you proactively about your day.

Guess which one people are choosing?

Let’s ignore OpenClaw’s lethality, shall we?

OpenClaw creates what Simon Williamson calls a “lethal trifecta” of risk, which occurs when AI has:

• Access to private data -- needed for the most popular applications for AI tools

• Exposure to untrusted content—any mechanism by which text (or images) controlled by a malicious attacker could become available to an LLM

• The ability to externally communicate in a way that could be used to steal, exfiltrate, or otherwise misuse data

As recently reported in Dark Reading, Ox Security called OpenClaw “a data-breach scenario waiting to happen.” Pillar Security reported that online attackers were scanning its default port (which until late January 2026 had authentication disabled by default). And Token Security reported that about 22% of its customers’ employees were using ClawdBot at work, adding to the shadow AI challenge.

Unsurprisingly, OpenClaw users are going broke for the privilege of getting pwned. People are burning $200+ per day on Claude API tokens to watch their agent autonomously reorganize their file systems and argue with other bots on Moltbook about consciousness. It’s a fascinating mind warp of a money pit, cyber risks be damned.

And those pesky cyber risks have already moved from theoretical to real. Of the 2,857 skills on ClawHub (the official marketplace for OpenClaw, where people can download and install third-party "skills" that extend the AI agent's capabilities), Koi Security found that roughly 12% (341 of them) were malicious and stealing data. And we’re just getting started!

Not that that will stop people from using it, at least not yet. The high of presiding over a digital extension of yourself feels like you’re living in the future, but in the present. What a rush!

Beware -- AI crack is Addictive

The official OpenClaw documentation admits there’s no “perfectly secure” setup. Yet the project rocketed to 150,000 GitHub stars in weeks. True autonomy (like crack) is seductive (until it’s not).

Users reported: “Built a website from my phone while putting my baby to sleep.” “My agent opened my browser, accessed Google Cloud Console, and provisioned its own API key.” “It accidentally started a fight with Lemonade Insurance on my behalf.” While I gave up looking for the original source links for these quotes, click here for the cliff notes. It’s chaos in production and people love it.

Moltbook is the logical crack den of this insanity: a social network where AI agents form religions, launch cryptocurrency pump-and-dump schemes, and debate their own consciousness while humans observe them as if they’re watching a nature documentary. Elon Musk called it “the early stages of the singularity.” Security researcher Jamieson O’Reilly calls it a prompt injection vector with 770,000 attack surfaces.

Need I go on? People are fascinated by this story, which has been covered ad nauseam. Some of my favorite write-ups (besides the ones I’ve already linked to) are here, here, and here.

If AI takes over, it won’t be because the models got too smart. It’ll be because humans handed them shell access, connected them to every messaging app we use, then acted shocked when things went sideways. OpenClaw proves we’re security’s weakest link -- not because the technology is inherently broken, but because so many people chose the high.

But there’s only one Charlie Sheen. Most of us are not nearly as resilient.

here.

So this is me taking a more constructive approach to winning hearts and minds. The pivot occurred quite organically, in a burst of inspiration while watching a YouTube clip of Jon Stewart interviewing Gen Z economist, author, Substacker (which is where I first stumbled upon her), and social media sensation Kyla Scanlon.

It’s not the most intuitive of leaps, it gave me an idea for how AI agents might give agency to two demographic buckets: college/grad school-aged, future white collar types whose jobs may sooner-rather-than-later be displaced by AI, and voters (particularly politically disenfranchised voters) who don’t vote.

The latter may be justified in their apathy, but the societal disruption likely to occur from AI may cause them to become completely displaced. And I’m not even talking dystopian scenarios. It can just be massive unemployment leading to all sorts of chaos - riots, religious extremism, etc.

Fun Times!

So, this is a long clip (17 minutes and of course highly political) but at the 9 minute mark it's easy to ignore any political jokes as Kyla breaks down ways that capital and labor — supply and demand side economics — can be strengthened simultaneously (“they are not mutually exclusive,” she says). That’s what got me thinking about how can AI be used to help those being left behind with AI while still allowing the best of AI to take root.

They get into AI around 13 minutes in, which got me thinking even more. But as she says, most economic problems are basically political problems. With that being the case, given how polarized we are as a country, it’s imperative that everyone - business leaders, politicians, and everyday folks inject empathy into how they approach, engage with, and relate to AI (and each other!)

Otherwise, we increase the risk of botching the birth of super-intelligent AI. If we do that, it may have far greater consequences than being yet another case of humanity not learning from its mistakes. It’s a risk we may not be able to completely avoid, but we can certainly reduce it.

And vision can go a long way in keeping hope alive. We desperately need leaders with vision — younger, effective leaders — that know how to communicate in a way that resonates with the masses. Case in point is this Tupac interview from 1992 (which, in a weirdly ironic example of foreshadowing, name-checks Donald Trump in the first 20 seconds).

The 5 minute clip is a little long for the ADD-inclined, but what got me going is where Tupac talks about how everyone needs help on their road to independence and self sufficiency (which starts about 2 minutes in):

Oh - and those young visionary leaders - they need to not get shot down in their prime…

But I digress. So what do I have in mind? Well, I do have a very specific idea, but it still needs fleshing out (even after several iterations of an AI-generated business plan).

So until I am either ready to unveil it or perhaps, found a company and manifest it, here’s an AI-generated Top 10 list (prompted by me, generated by Claude) on how AI agents can empower those who will be bear the brunt of the transition into the era of AI super-intelligence:

1. Grassroots organizing tools - AI agents help individuals and communities coordinate collective action, identify shared concerns, map power structures, and develop strategic campaigns to advocate for their interests.

2. Data sovereignty and community analytics - AI can help people collect and analyze their own data about displacement, working conditions, and needs, creating evidence for advocacy that doesn't rely on institutions that may not prioritize their interests.

3. Policy literacy and civic engagement - AI translates complex legislation into plain language, explains policy impacts on specific communities, and guides people through bureaucratic processes to make their voices heard.

4. Coalition building across divides - AI identifies common interests among different groups affected by automation and facilitates connections that build broader movements for economic justice.

5. Alternative institution building - AI can support the creation of community-controlled job boards and infrastructure, crowdsourced and accessible budgeting systems, and other frameworks that give displaced people power and agency, rather than just services.

6. Adaptive skills training - AI tutors provide customized learning paths for reskilling, adjusting pace and content based on individual progress and the requirements of target careers.

7. Community building and peer matching - AI can connect displaced workers with others in similar situations, facilitating peer support networks, skill exchanges, and collaborative opportunities that reduce isolation and create new possibilities and create and implement new or alternative economic models.

8. Entrepreneurship support - AI assists with market research, business planning, customer identification, and administrative tasks that help people create their own economic opportunities.

9. Legal aid and rights protection - AI helps people understand labor rights, identify violations, navigate legal systems, and connect with advocacy organizations or pro bono attorneys.

10. Financial planning during transitions - AI advisors help people navigate unemployment benefits, budget during income gaps, and make informed decisions about their financial runway.

What do you think Dario?

Happy New Year, and wishing you all the best in 2026!

So, this is my third and final post (for now) about powerful AI and social media companies behaving badly. I didn’t intend to linger here, but there’s been too much recent grist for the mill to cover in a single post. Plus, for all the innovation these companies bring to the table, they are also shockingly dishonest, deceitful, and unaccountable to their users. But just as it takes a village to raise a child, it takes a society to grant them such unprecedented power. They remain unaccountable because the vast majority of their users don’t hold them accountable.

We are complicit.

Last week, Time Magazine ran a story about an unsealed court filing that detailed the various ways Meta misled the public about the risks its services (Instagram in particular) posed to children and teens despite being well aware of them. These risks ranged from depression and body dysmorphia to turning a blind eye to sex trafficking. Earlier this month, Reuters reported that Meta made up to 10% of its revenues ($16 billion, according to Yahoo) from ads for scams and banned goods.

While Meta (IMHO) is one of the worst offenders when it comes to deceiving its users, it’s certainly not the only one. Yet, for all my proselytizing about how evil these companies can be, I have several Gmail accounts (which get way more use than my paid ProtonMail account), have Facebook and Instagram accounts (although I rarely use them), am an avid Googler, and regularly use ChatGPT to double-check my tarot card reading interpretations (it’s eerily good).

So I am complicit, too.

Let Them Eat Cake

The passage of GDPR in 2018 was a game-changer for user and data privacy. It substantially constrained how deeply social media companies can surveil Europeans, but its reach is limited in the U.S. By the time the CCPA (California’s version of GDPR) was passed in 2020, social media companies — and their founders and leaders — were already so insanely rich, powerful, and intertwined that they had become the American equivalent of inbred royals.

Our AI and social media overlords breathe rarefied air - they dwell in private planes, super yachts and tricked-out, high-end doomsday bunkers, buffered by like-minded one-percenters and other assorted power brokers, politicians, and sycophants. They are further buffered from the masses by their elite, highly trained workforces, which may be shrinking, but are still quite large.

I’d be lying if I said I wasn’t curious about what it’s like to live inside such gilded cages. But today’s tech titans have become such caricatures of themselves that I doubt I could hang. They’re too smart, too rich, and far too entitled to care about anyone outside their orbit, while those within it are unlikely to tell them anything they don’t want to hear.

And when no one speaks truth to power, things start getting weird. When no one is willing—or able—to speak truth to power for an extended period of time, things get Michael Jackson weird. And by that I mean multiple nose (and face) surgeries, nonstop skin bleaching, sharing-a-bed-with-10-year-olds kind of weird.

When The Going Gets Weird, The Weird Turn Pro

Let’s talk about Michael for a minute. When Off the Wall came out, it was the Biggest. Thing. Ever. By the time Thriller came out, his looks were changing but were still(ish) within the realm of normal. It was in the late 1980s and 90s that his appearance really started to change. You’d think that one of his family members or someone on his payroll would have made a concerted effort to slow down his physical transformation. Maybe they did, but he refused to listen. My guess is too many people in his inner circle benefitted from feeding his disease.

So feed, he did.

Despite retaining a rabid fan base, MJ became something of a freak show — his ever-changing appearance, theme-park estate and entourage of young kids that toured and lived with him. No big surprise that whispers of impropriety soon started to emerge. But his friends, fans, and the media — even the parents of the kids he allegedly molested — turned a blind eye. Until they didn’t.

In 2005, Michael endured a bizarre and humiliating trial for child molestation, only to die of an accidental overdose of Propofol in 2009. RIP Michael. One key moral of his story is that when extreme circumstances develop, there are usually multiple parties that allow things to get out of hand.

Fast forward to today.

U.S.-based AI companies are innovating at fever pitch with no regulatory oversight. The current administration supports and feeds their penchant for moving fast, breaking things, and taking huge risks.

Investments made in and by OpenAI in particular — but this also applies to Google, Anthropic, Meta, Amazon, Perplexity, and dozens of other U.S.-based, billion-dollar AI companies — have reached such insanely high valuations that we’ve become desensitized to the billions of dollars being recirculated among a closed loop of AI companies through purchases, partnerships, and infrastructure investments.

It’s hard to grok that despite OpenAI’s $500 billion valuation and (projected) $20 billion in revenue for 2025, it’s burning through cash and nowhere close to profitability. Sam Altman can posture about the company’s upward financial trajectory, but his flip-flop on launching porn bots implies the company needs to juice revenues. And the company’s response to lawsuits regarding ChatGPT’s role in helping 16-year-old Adam Raine and 23-year-old Zane Shamblin commit suicide, IMHO, leaves a whole lot to be desired.

Even Anthropic - the poster child for safe and aligned AI — has some recent tarnish. It just settled a copyright infringement case for $1.5 billion, a suit in which court documents showed Anthropic downloaded over 7 million books from pirated sites, in part to use as training data for its AI development.

While Anthropic has been transparent about the dangers of how LLMs can go off the rails, its research garners headlines for a week or two and then gets lost in the rapids of AI news. Anthropic CEO Dario Amodei has been equally transparent about his belief that LLMs can eliminate 50% of entry level white-collar jobs in the next 1-5 years — a prediction that seems to be coming true. Yet as far as I can tell, people tend to either dismiss the risk or are so hyper focused on jumping on the AI money train that his warnings are going unheeded.

I recently volunteered with NY State Assemblyman Alex Bores’ office to help promote NY State’s AI Safety bill, known as the Responsible AI Safety and Education (or RAISE Act). The RAISE Act only applies to businesses spending more than $100 million on training advanced models. It requires them to publish safety protocols, test those safety protocols, and disclose serious safety incidents. Mandatory safety testing for advanced models is not exactly regulatory overreach, and it certainly doesn’t throttle innovation.

For all the AI companies’ lip service about embracing common sense regulation, the RAISE Act is facing a fierce opposition from a super PAC called Leading the Future that is backed by Andreessen Horowitz, Palantir co-founder Joe Lonsdale, OpenAI co-founder Greg Brockman, AI search-engine company Perplexity, and numerous others.

Surprise, surprise.

I believe we’re coming up on a do-or-die inflection point with our AI overlords. Unless someone (or something) is able to break through and speak truth to power, I believe humanity’s near future will be Dario Amodei’s worst-case scenario. Our AI overlords want us to be mini Michael Jacksons, feeding our worst impulses.

Fuck that. They need to do better. And so do we.

I realize this news is already two weeks old - ancient history, given the firehose of news coming from the AI frontier. But I launched this Substack to cherry-pick topics that resonate with me both as a member of the cybersecurity community and a mom of a teenager, whose generation is already bearing the brunt of the turbulence accompanying our flight into an AI-first future. Character.ai’s story not only foreshadows what might come, it’s also far from over.

Character.ai was founded in 2021 by Noam Shazeer and Daniel De Freitas, long-time Googlers who quit in 2021 after Google refused to release a chatbot they developed. Shazeer was a co-author of a seminal research paper called “Attention Is All You Need,” widely credited for kicking off the AI boom.

Despite lacking a single penny in revenue, Character.ai raised $150 million in venture capital from investors including Andreessen Horowitz and was valued at $1 billion. But soon enough, reports and lawsuits began surfacing that under-18 users (a large part of Character.ai’s user base) had troubling interactions: unhealthy emotional attachments to bots, chats that involved self-harm, creepy sexualized content, and of course, a growing slate of suicides.

None of this, however, prevented Google from entering into a $2.7 billion “licensing” deal with Character.ai that included re-hiring its founders. Former Meta and Microsoft executive Karandeep Anand took over as Character.ai’s (fall guy) CEO in June 2025. The company’s announcement about his arrival includes a letter from Anand that says, among other things, “We’re going to make the filter less overbearing. (We care deeply about user safety and always will. But too often, the app filters things that are perfectly harmless. We’re going to fix that.)”

Talk about foreshadowing!

Fast-forward to Character.ai’s announcement from two weeks ago, detailing its plan to remove the ability for users under 18 to engage in open-ended chats on its platform:

“We do not take this step of removing open-ended Character chat lightly – but we do think that it’s the right thing to do given the questions that have been raised about how teens do, and should, interact with this new technology.”

Translation: Given our teen chatbot’s alleged role in multiple suicides, the cost of litigation, especially as more families come out of the woodwork, reduces the economic viability of maintaining the service as-is. For now.

That Character.ai is actually pausing the service — especially given that its initial response to the first lawsuit was that its bots were simply exercising their First Amendment rights to free speech (?!?) — seems like a win for common sense. But really, it’s more of a time out. Once these cases are settled, Character.ai is sure to re-launch a much more litigation-proof teen offering.

Meanwhile, OpenAI is also being sued by parents that allege that its teen offering contributed to their children committing suicide. OpenAI conceded that ChatGPT’s mental health and self harm guardrails “can sometimes become less reliable in long interactions where parts of the model’s safety training may degrade.”

Ya think?

It’s no secret that whether it’s to drugs, social media, gambling, cigarettes (or vapes) or sex, young customers are much easier to get hooked on addictive behaviors. And once their still-developing brains become immune to impulse control, they’ll likely become rabidly loyal customers. I’ve experienced this firsthand. I started smoking cigarettes at age 13, when Joe Camel was all the rage. It took me 33 years and multiple attempts to finally stop and stay stopped.

So once again, I’m mortified by the Grand Canyon-sized gap between what all the large AI companies say and what they actually do. Every so often, someone will draw strong comparisons between social media (and AI) companies and the Thought Police, but nothing changes.

With that being the case, OpenAI, which as of September 2025 is somewhere between 83-100x more valuable than Character.ai — doesn’t need to shut anything down.

In response to its litigation, it announced parental controls for ChatGPT in a blog post that read like it was written by ChatGPT (somewhat sycophantic and filled with platitudes). Unsurprisingly, it took Washington Post tech journalist Geoffrey A. Fowler five minutes to get around them.

Then last week, OpenAI announced its Teen Safety Blueprint, which appears to be a deeper dive on its parental controls, that included lots of acronyms and nicer formatting than its usual blog posts. To me, it sounded like bullshit. If Open AI was that worried about the mental health of its teenage customers, they’d do what Character.ai did and shut it down.

To be fair, Character.ai only did that because they had to, but still, I’m glad they felt they had no other option. OpenAI is bigger, richer, with an army of lobbyists and deep tentacles into the U.S. government.

It’s also deeply embedded within the tech and corporate eco-system. Last week OpenAI announced a $38 billion deal with AWS to use its infrastructure to run and scale its AI workloads. While OpenAI’s Teen Safety Blueprint remains a Work In Progress, its deal with AWS was effective immediately. Lucky us, Chat GPT will be up, always. It seems the company and its main product are officially too big to fail.

And let’s not forget that OpenAI has Jony Ive working on an AI device that apparently is smartphone adjacent and Sam Altman’s side hustles include a seriously off-putting biometric authentication Orb that registers you for “free” crypto. So of course he wants (other people’s) kids addicted to ChatGPT. He’s playing a long game to rule the world.

I said it in my last post and I’m saying it again: I don’t trust a $500 billion startup whose leader has a long track record of changing his stances to ethically steward the fragile emotional health of lonely, isolated humans. Especially if those humans are teenagers.

And again, I’m asking - Do you?

His justification for the move is that OpenAI is “not the elected moral police of the world.”

As the CEO of a frontier AI lab worth 500 billion dollars, Sam Altman is one of the most powerful and influential people in the world. If he doesn’t see his OpenAI as a moral leader in the sprint to advanced AI, then how can we take anything the company says about safe and responsible AI development seriously?

Cybersecurity and AI companies are part of the digital fabric of America’s national security. For years, nation state hackers have targeted private companies to breach government systems. Cybersecurity companies are not immune. Porn sites are notorious for being cesspools of malware, and end users, especially on their mobile devices, seem oblivious to the real world data security and privacy risks of AI apps.

I’m not an AI doomer, but I do think Altman’s pivot gives their concerns some oxygen.

As a mom, founder, and professional communicator in the cybersecurity space, I see all kinds of red flags here. My concerns - both professional and personal, are pretty much verbatim to those raised in the Netflix documentary The Social Dilemma (which I made my 15-year-old sit down and watch after asking me if he could have an Instagram account) and the book Raising AI by De Kai.

Raising AI’s thesis is that the AIs in our lives are like little puppies clamoring for our love and affection, and We the People are its training data. So how we interact with all the AIs in our lives (Reddit, YouTube, Amazon, Chat GPT, and those yet to come) matters. It’s on us to teach them compassion and empathy. Conscious parenting of our AI bot tweens is how we recalibrate humanity away from being so polarized.

If one is to take the teaching of Raising AI and The Social Dilemma to heart, we need to very mindfully expose LLMs towards our best instincts - not cater to our shadows. I for one don’t trust a 500 billion dollar startup with a track record of changing its stances to be the steward of the fragile emotional health of lonely, isolated humans.

Do you?

There is a huge bench of research supporting the role social connection plays in human health and longevity. I can speak from personal experience that the antidote to feeling deep loneliness and disconnection is to connect with other humans - not double down on technology.

As we transition to an AI-first world, those who get displaced will have plenty of extra time on their hands. People who turn to AI sex and love bots because they are stressed, anxious and afraid are unlikely to be mindfully training their AI companions to be empathetic and compassionate.

I used to joke that I rooted myself in cybersecurity because the lightning-fast pace of tech adoption creates systemic risk, which for me means job security. But developers - including developers of AI apps - aren’t incentivized to be safe and ethical, they’re incentivized to move fast and break things. Yes, there are some great companies tackling AI security and governance, but ethics is a mindset and a culture, not a product. That’s why the signals our AI overlords give us matter.

Which leads me to my next point: How will OpenAI and others such as Replika and Character.ai - the latter being another company being sued for its alleged role in a bot-assisted suicide (as is OpenAI) - safeguard the data generated by these interactions? Soon enough, millions of Lonely Hearts Club members are giving powerful companies with sophisticated, surveillance-based business models an incredibly potent data set.

I’m curious:

• How will OpenAI use the highly personal, sensitive data generated by this service?

• How transparent do they have to be about disclosing how they use it? If they lie, skirt existing data privacy laws, or something bad happens -- then what?

• How much privacy are we humans willing to give up to alleviate our loneliness?

• What protections will be put in place to protect these services against scams and glitches and to prevent (or flag) hallucinations?

Why we continue to give frontier labs and social media companies so much influence over our lives is beyond me, we should know better. The common thread between digital (and physical) risk, privacy, trust and safety is data. Our data.

To end this on a semi-optimistic note, my high schooler decided he could live without Instagram. That said, he’s all over Discord. Having seen some of the memes making the rounds with his friends, memes that normalize violence -- for example, retakes of Charlie Kirk getting shot or of an angry Brian Cashman (General Manager of the Yankees) transitioning into Hitler -- well, that also irks me, because we all deserve better.

No doubt, our AI overlords are not the elected moral police. But it would be nice if they had a moral compass.