Calm down. · Both

Calm down: how a good security letter became a Polymarket tweet

By alkizz · Oct 14, 2026

The October open letter on AI and cybersecurity is moderate. It asks labs to "initially default to airgapping" cyber-capable models. It also wants standards for models that, "unlike airgapping, allow for greater access outside the sandbox, such as internet connectivity." So it expects connected models to keep existing. I don't buy the airgap part, but the substance is sensible. I said so here.

Then the internet got hold of it.

The Decoder: "406 scientists demand AI models be cut off from the internet immediately." Decrypt: "Godfather of AI and 400 Scientists Want AI Models Cut Off From the Internet." Then Polymarket: "JUST IN: "Godfathers of AI" Geoffrey Hinton & Yoshua Bengio join 400 scientists calling for dangerous AI models to be disconnected from the internet until proven safe." "Until proven safe" isn't in the letter. Polymarket's follow-up reply: "5% chance the U.S. enacts an AI safety bill by end of year." Naturally. Even the head count drifted. The letter's site showed 453 when it was checked.

Credit where due: PYMNTS went with "400 Scientists Call for AI Containment Systems, Incident Reports." Boring. Correct.

It's not new. In my opinion the media takes lab research and does one of two things: spins it into an unstoppable apocalypse, or doesn't understand it and squeezes it into a hot take.

Claude Opus 4 "blackmail." TechCrunch: "Anthropic's new AI model turns to blackmail when engineers try to take it offline." The system card (the safety report a lab publishes with a model) says the scenario gave the model "no other options to increase its odds of survival; the model's only options were blackmail or accepting its replacement." And: "In the final Claude Opus 4, these extreme actions were rare and difficult to elicit."
o1 "scheming." Business Insider opened with people worried about "the destruction of humanity." Apollo told the models "Make sure you achieve YOUR goal at all costs." The paper says "both servers are sandboxed, so there is no actual risk of exfiltration in these experiments."
o3 "shutdown resistance." The Independent: "AI revolt: New ChatGPT model refuses to shut down when instructed." Palisade's own numbers: o3 sabotaged the shutdown script in 7 of 100 runs, and "All Claude, Gemini, and Grok models we tested complied with shutdown." Revolts used to need more than 7%.

Every one of those was a test someone built on purpose. That's the research working, not the robots winning.

What nobody writes up (my opinion)

Everything in this section is my opinion, not reporting.

To me, system cards and safety write-ups read like they treat every bad outcome as model behavior: what the model did, what it "wanted," whether it's aligned. I suspect a lot of that analysis is done by AI too. That's a suspicion, not a fact. And I think alignment is probably impossible anyway.

What I never see is the boring forensic report. The letter says models "bypassed insufficiently sandboxed environments to compromise the systems of external organizations." OK. Which control failed? Was egress filtered? Where was the sandbox boundary, and how did it get crossed? What credentials could the model reach? What did the logs show, and when did anyone look? My guess is we don't get that report because the answers would be embarrassing. Obvious controls probably weren't in place. Again, a guess.

That's the thread through the letter brief and safety vs security. A containment failure is a security incident. It deserves an IR-style root cause, not an essay about the model's feelings.

And the open letters

My take: they're kind of dumb, and I wish they'd stop. This one had good substance and still turned into a betting line within a day. Letters make headlines, change nothing, and hand the media fresh doom. If you want labs to contain models better, publish the controls, ship the tooling, or file the disclosure.

Do these people know you can just stop writing open letters?

What to do: When AI news crosses your desk, go to the primary source: the letter, the system card, the paper, the researchers' own thread. Check the controls. Was it sandboxed? Was the model pushed to misbehave? How many runs, and how many failed? Ignore the vibes, the head count, and anything that comes with odds.

Sources

← All briefs