AI Brief, October 21: eMudhra CA compromise, Anthropic's safety bill, TorchTPU

Compromised eMudhra intermediate CA keys force trust store updates, Anthropic drafts a federal AI safety bill, and Google ships TorchTPU.

The day's most urgent item is operational, not model-related: private keys for eight eMudhra intermediate CA certificates have been compromised, and anything bundling a trust store needs updating. Policy also moved, with Anthropic circulating a draft federal AI safety bill and the US ordering DeepSeek out of federal civilian systems, while Google and Amazon pushed PyTorch onto their own silicon.

Eight eMudhra intermediate CA private keys compromised, trust stores need updating

Simon Willison flags the compromise of private keys for eight eMudhra intermediate CA certificates, which chain to roots trusted by the Mozilla, Microsoft, Apple and Chrome root programs and many downstream trust stores. Any software that bundles a list of trusted certificates needs to be updated, and Linux distributions need updated ca-certificates packages. In the Python ecosystem, uv 0.13.4, pip 26.2.2 and certifi 2026.10.19 have shipped with upgraded certificate lists. A GitHub repository is tracking software updates.

Why it matters: If your build images, CI runners, or shipped binaries pin an old CA bundle, they can silently trust forged certificates until you update them.

Sources: Simon Willison

Anthropic circulates draft federal AI safety bill with catastrophic risk thresholds

TechCrunch, citing Politico, reports that Anthropic has been shopping a draft federal AI safety bill that would allow a model release to be blocked when there is a "foreseeable, imminent risk" to more than 50 lives or large-scale property damage stemming from a biological weapon, a model evading developer controls, or a cyberattack. Independent evaluators registered with the Commerce Department would have 72 hours to report such a risk, and Commerce would then have to ask a Washington, D.C. judge to halt the release. The bill would expressly forbid any federal agency from halting distribution on its own. An Anthropic spokesperson confirmed the draft to Politico, saying it reflects the company's Advanced AI Framework.

Why it matters: A registered-evaluator regime with court-ordered release halts would change pre-launch compliance work for anyone training or distributing frontier models in the US.

Sources: TechCrunch AI

US orders federal civilian agencies to remove DeepSeek and products built on it

TechCrunch reports that the US has ordered federal civilian agencies to stop purchasing DeepSeek products and services and to remove them from government IT systems, according to the Department of Homeland Security. The order covers third-party products and services built using DeepSeek technology, though agencies can request exceptions. DHS cited concerns that sensitive US government user data could be accessible to Chinese authorities under Chinese law. It extends a December 2025 law that targeted DeepSeek's use in national security systems; China's Foreign Ministry spokesperson Mao Ning said Beijing opposes "overstretching the concept of national security."

Why it matters: If you sell into government or regulated buyers, open-weight DeepSeek models anywhere in your stack are now a procurement blocker, so check your model provenance.

Sources: TechCrunch AI

Google open sources TorchTPU a day after Amazon's TorchNeuron

TechCrunch reports that Google released TorchTPU, its version of PyTorch for Tensor Processing Units, as an open source project on Tuesday, saying existing PyTorch apps can move to TPUs with a configuration change and no rewriting. Google says it worked with Meta, Hugging Face, Inferact (vLLM), RadixArk (SGLang) and others, and launched a TPU agent suite to help with migration. Amazon released TorchNeuron for its Trainium chips on Monday and is submitting code to the main PyTorch project to decouple features from Nvidia hardware. Hugging Face says TorchTPU adds TPU as a native PyTorch backend, so Transformers and Diffusers models can be loaded with device_map="tpu" and Optimum TPU is no longer required.

Why it matters: Porting cost has been the main reason to stay on Nvidia, and two major clouds just reduced it to a config change for standard PyTorch workloads.

Sources: TechCrunch AI / Hugging Face

Anthropic says Claude designed 17 validated protein binders against scorpion neurotoxins

Anthropic says an internal research model, working within Claude Science from one prompt plus four follow-ups over 64 hours, ran a de novo design campaign against five scorpion neurotoxins from four regions. Anthropic says Claude generated 318,484 designs using 12 design tools, screened 553,052 design-toxin pairs, and selected 180 for testing, with 31 more added by the team. Adaptyv Bio and Twist Bioscience lab-tested all 211 designs by surface plasmon resonance, and 17 met Anthropic's binder criteria, each against a single toxin; none of the multitoxin designs qualified for more than one. Anthropic says it has not confirmed neutralization and does not claim to have produced an antivenom, and is releasing all 211 designs and their data.

Why it matters: It is a documented case of a long-horizon agentic run producing lab-verified results, with the company's own caveats about what was and was not validated.

Sources: anthropic.com / Hacker News discussion

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.