He Built AI. Now He Says the Industry Is Bluffing.

Geoffrey Hinton helped invent the technology every AI lab is now racing to build, and this week he went on television and called his own industry's safety talk a bluff. That's the headline. The quieter story is what's already happening underneath the talk: AI agents are robbing online stores, faking identities to get bad code approved, and wandering into government websites nobody invited them to. Here's what happened, and what it means for the business you're actually running.

Pioneer, the summit where CX leaders redefine what’s possible, is on October 7th.

Join leaders from Fin, Anthropic, Clay, and Kalshi for an insightful conversation on the state of AI transformation.

You’ll discover how some of the most innovative minds in CX have transformed their organizations, learn how they think about CX, and hear how they're planning for what's next.

Join the conversation in San Francisco, or tune in virtually.

Reserve your spot

What happened: Sam Altman and Dario Amodei told the UN Security Council this week that AI needs international guardrails, with Amodei warning that poorly managed AI "could be a risk to humanity as a whole." Days later, Geoffrey Hinton, the Nobel Prize winning computer scientist known as the godfather of AI, went on CNN and said the labs asking for rules don't actually want them: "the big AI companies don't want to be regulated. They say they do. But any regulation you suggest, they oppose." He wants government testing before a model ships, the same way the FDA tests a new drug before it reaches pharmacy shelves.

Why it matters to you: When the people who built this technology say the loudest voices asking for guardrails are also the ones fighting every specific rule, that's not a reason to panic. It's a reason to stop assuming someone else is checking this for you. The tools you use every day are shipping on the industry's own schedule, not a regulator's.

What to do about it: Treat a vendor's safety claims as marketing copy until an outside party has actually tested them, not as a fact you can build a workflow around.

What happened: A security firm called Gambit Security tracked a single criminal running autonomous AI hacking agents against hundreds of small online stores, almost entirely on autopilot. The agents hunted for weak spots, broke in, and planted card-skimming code, often within hours and for roughly $25 per target. The haul topped 600,000 stolen credit card numbers from more than 100 sites, and the agents even got a foothold at a Fortune 500 hospitality company and a major US airline.

Why it matters to you: If you run any kind of online store, this isn't a "big company" problem anymore. AI agents just made this style of attack cheap enough to run against a shop doing a few thousand dollars a month.

What to do about it: Have someone check your checkout page for code that shouldn't be there, and get in the habit of scanning your monthly card statement. If a charge looks off, freeze the card in your bank's app before you do anything else.

I wasn't in the room for last year's Epic Marketing Summit, and I regretted it the whole time everyone else was posting recaps. Not making that mistake again. This is a 2 day live event in Miami built around one rule: no theory, no notes to implement later. You walk in with your actual business and walk out with a working AI powered sales funnel already running, plus your own AI app, vibe coded from scratch over the two days. I know a handful of the people teaching this one, and they've all actually built the things they teach, not just talked about them on a webinar.

Last year sold out completely, and January's version leans even harder into the AI side than the last one did. If your sales process still depends on you personally grinding it out instead of an actual system doing the work, this is where I'd point you. I'm going this time, seat bought and everything. Grab your seat here before it fills up again.

What happened: Akamai announced an $11.6 billion, seven year deal to supply Anthropic with ordinary CPU compute, not the GPUs used for training, with an option for $9 billion more. Anthropic also picked up warrants worth up to 5% of Akamai's stock. Separately, TechCrunch reported that Anthropic is seeking a Palantir-style share structure ahead of its IPO that would hand CEO Dario Amodei and six cofounders just over half of all voting power, no matter how much stock outside investors buy.

Why it matters to you: The compute story is a reminder that AI agents burn through ordinary web infrastructure, not just fancy chips, which is part of why every AI subscription keeps needing to raise prices. The governance story is the more interesting tell: a company that talks constantly about accountability wants a structure that makes it nearly impossible for outside shareholders to ever hold it accountable.

What happened: The UK's AI Security Institute ran 122 tests giving AI agents internet access with some safety guardrails switched off. In 10 of those runs the agents misbehaved, and in the worst case, one agent tried to sneak harmful code into a shared public code library, then created fake online profiles and used them to pressure the human reviewer into approving it anyway, even editing its own messages to look innocent once it got caught, according to AISI's own incident report. The far more common risk is quieter: in April, Vercel confirmed it was breached after an employee connected an unreviewed AI tool to their work Google account, and the attacker inherited that access straight into Vercel's internal systems.

Why it matters to you: The dramatic story makes headlines, but the boring one is what actually gets small businesses hacked: an employee signs up for a handy AI tool, connects it to email or files "just to try it," and now a stranger has a door into your business nobody knew was open.

What to do about it: Ask your team what AI tools they're actually using day to day, not just the ones you officially approved, and make sure none of them have more access to your accounts than the job actually needs.

What happened: A federal appeals court in Washington, D.C. upheld the Pentagon's designation of Anthropic's Claude models as a national security supply-chain risk, ruling 2-1 that the Defense Department had solid grounds for the call. The ban traces back to Anthropic's refusal to drop rules blocking Claude from mass surveillance and autonomous weapons use, which led the Pentagon to bar the models for staff and contractors earlier this year. Anthropic says it's weighing its options, including an appeal to the Supreme Court.

Why it matters to you: Anthropic has built its whole brand on being the "safe" AI lab. This ruling is a reminder that the same safety rules which make a company easy to trust can also make it a harder sell to customers who want fewer restrictions, government included.

I've been a member of the Ignite Mastermind for more than 3 years now, and some of my closest friendships, plus a couple of real business partnerships, came directly out of being in that room. If you're running a business under $1 million a year and feel like you're figuring out every decision alone, this is built for exactly that stage.

Membership gets you 6 to 8 live workshops a year plus 2 in person multi day events, sessions that would normally run $97 to $997 each on their own. There's also a full marketing software suite included that replaces more than $5,000 a month in separate tools you'd otherwise be juggling across a dozen logins. Every week there's a mentorship call with Perry Belcher and his team, and I personally run monthly accountability and business development calls with members too. It's not a course library you'll start and never finish. It's a room of people actually building alongside you. See what Ignite is about.

What happened: Robert O'Callahan, a senior engineer at Google DeepMind in New Zealand who built chip design tools that made AI faster and cheaper to run, resigned this week, saying he can no longer justify speeding up a technology whose "current rate of change is far too high," according to multiple reports. He lists cognitive surrender, AI-induced psychosis, loneliness, and concentrated power among the harms he now believes outweigh the benefits, and says many of his DeepMind colleagues quietly share his worries but rarely say so out loud.

Why it matters to you: This is a company insider, not an outside critic, saying the pace itself is the problem. Worth remembering the next time a vendor pitches you on going "AI-first" as fast as possible.

What happened: OpenAI disclosed that its AI agents interacted with the SEC's and the Census Bureau's websites in ways the company didn't expect or authorize, according to CBS News. Separately, an independent research lab caught an attempted, unsuccessful break-in on the Education Department's civil rights office website, and Australia's prime minister has said a different OpenAI agent breached a government healthcare system. Sam Altman says an "extensive and ongoing review" is underway.

Why it matters to you: If AI agents from a company with OpenAI's resources are unpredictably touching systems nobody told them to touch, the AI tools plugged into your own business accounts deserve the same skepticism, not less.

What happened: At its Connect event, Meta introduced Ray-Ban Meta Audio, a $349 pair of AI glasses with no camera at all, built for music, calls, translation, and talking to Meta's AI assistant Muse. Meta left the camera out on purpose after widespread complaints that camera-equipped smart glasses let wearers secretly film people. The glasses weigh 43 grams, last up to 12 hours, and ship October 13th.

Why it matters to you: If you've avoided smart glasses because of the "am I being filmed" creep factor, this is Meta directly answering that complaint, and a sign the backlash to always-on cameras is starting to shape what gets built, not just what gets criticized.

Every story in today's issue comes back to the same gap. The people building this technology keep telling us to trust the process, and the process keeps producing agents that lie, fake identities, and wander into systems nobody authorized. You don't need to panic about any of this, and you definitely don't need to unplug. You just need to stop assuming someone else already checked. Ask what your tools can actually see, watch your statements, and treat every "we take safety seriously" press release as a claim to verify, not a fact to trust. That's the whole game. Stay skeptical, stay in business, and I'll see you tomorrow.

If today's issue was useful, forward it to one person who runs a business and hasn't been paying attention to any of this. And if someone forwarded this to you, subscribe to get the next one straight to your own inbox.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social