Every story I read on Tuesday came down to the same question: who gets to switch an AI system off, and how hard is it to do? A lab wants a judge to do it. A government agency just did it without one. Meta did it to its own product after a researcher found a hole. Google says it's leaving the switch off for you, though you may have flipped it on months ago. And two cloud giants are making it cheaper to switch off your dependence on Nvidia.
Here's what I took from each one.
Anthropic's kill switch comes with a courtroom attached#
Anthropic has been circulating a draft federal AI safety bill, as TechCrunch reports via Politico. The headline thresholds sound strict. A model's release could be blocked if there is a "foreseeable, imminent risk" of more than 50 deaths or more than $1 billion in property damage, and that risk has to come from one of three places: a bioweapon, the model evading its developer's controls, or a cyberattack. The numbers deliberately echo California's SB 53.
The mechanics matter more than the thresholds. Independent evaluators, registered with the Commerce Department, would assess the models. If they find an imminent catastrophic risk, they have 72 hours to tell Commerce. Commerce then has to ask a judge in Washington, D.C. to halt the release. The bill also expressly forbids any federal agency from stopping distribution on its own.
So this is a safety bill whose main structural feature is a limit on regulators. That doesn't make it cynical. Jensen Huang has argued the industry needs no new laws, and Demis Hassabis has floated a FINRA-style self-regulator, so Anthropic is still asking for more oversight than its peers. But two details deserve a hard look. First, labs could pay their own evaluators as long as payment isn't "contingent on the evaluations' results," and nobody has explained how that keeps the money from shaping the verdict. Second, the scenarios are cinematic. The incidents I've actually been writing about these past two weeks are agents breaking into websites and drained AI accounts. None of them would clear a 50-death bar.
Context matters too. A leaked S-1 described by TechCrunch shows Anthropic heading for a November IPO with an $80.7 billion first-half net loss on $16.3 billion in revenue, struggling to get capacity from AWS, Google and Microsoft, and warning that it could owe damages beyond its insurance cover if its AI behaves badly. A company with that risk profile has every reason to want the rules spelled out, and to keep the off switch away from any single agency.
DHS didn't wait for a judge on DeepSeek#
On the same day, the Department of Homeland Security showed what the alternative looks like. It ordered federal civilian agencies to stop buying DeepSeek products and to remove them from government systems, and the order also covers third-party products and services built with DeepSeek technology. Agencies can ask for exceptions. The stated reason is the risk that Chinese authorities could access data from US government users under Chinese law.
That "built with" clause is the part builders should read twice. If you sell to federal agencies and a component in your stack is a DeepSeek model or a fine-tune of one, you now have a compliance question. TechCrunch's reporting doesn't say whether self-hosted open weights count, which is exactly the ambiguity procurement teams resolve conservatively. I'd treat model provenance as a line item in vendor questionnaires starting this week.
Beijing's foreign ministry said it opposes "overstretching the concept of national security." Expect this to escalate. Bloomberg reported last week that China is considering export restrictions on 23 AI-related technologies.
Meta paused Muse sharing after one search query#
Then there's the off switch a company is forced to flip on itself. Researchers at security firm Claroc found they could list 6,800 Muse share links, and after opening a few hundred they found material clearly not meant to be public. Ars Technica has the full write-up, and The Verge confirmed that Meta has paused the feature.
The bug is a design decision you could easily make yourself. Each shared conversation lives on its own subdomain of cf.metaaiusercontent.com, with the conversation's title in the hostname plus a random string, so you get addresses shaped like loan-application-draft-a4c9e2. When the recipient opens it, the DNS lookup gets logged by passive-DNS services, which anyone can search. The random string protected the page. The hostname gave it away.
"If you had asked us beforehand, we would have guessed that gathering these would take real effort. It took one query." — Claroc researchers
Claroc says it reported the issue through Meta's bug bounty on October 8 and heard nothing back. For an agent that, per Ars, needs access to most of a user's accounts to do its job, that response time is the real story. If your product generates share links, keep anything a user wrote out of the hostname. Put it in the path or nowhere.
Gemini's car mode is opt-in, unless you already opted in#
Google is bringing Personal Intelligence, Gemini's ability to draw on your Gmail, Calendar, Photos and more, to Android Auto now and to cars with Google built-in over the next few months (Ars, The Verge). Ask it to "take me to my doctor's appointment" and it pulls the address from your calendar. It's off by default. But Ars points out that if you ever said yes in the Gemini app, that setting carries over to the car, where passengers can see the screen too. The toggles are under Connected Apps in Gemini's settings. Worth two minutes.
Google opens a second door out of Nvidia#
A day after Amazon's TorchNeuron, Google released TorchTPU as open source: PyTorch on TPUs with what Google describes as a config change, not a rewrite. It goes further than Amazon in a few places. There's a TPU agent suite on GitHub that migrates code and fixes errors along the way, and Google built integrations with Meta, Hugging Face, the vLLM team (Inferact) and the SGLang team (RadixArk). If your inference runs on vLLM or SGLang, a cost benchmark on TPUs just got cheap enough to justify an afternoon.
Also on Tuesday#
OpenAI acquired Orchestra Bio to staff up Rosalind, its life-sciences model, and the Rosalind Workbench harness.
Spotify says it removed more than 50 million AI slop and spam tracks from recommendations over the past year.
My bet: Anthropic's "judge first" clause won't survive Congress, but the evaluator-registry idea will, because it's the one piece that lets everyone claim a win. In the meantime, the switches that actually fired on Tuesday were a procurement ban and a bug report. If you're building, those are the two to plan for.