Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- Physicist Christopher Stubbs, former Harvard Dean of Science, joins Anthropic's Long-Term Benefit Trust
Anthropic says experimental physicist Christopher Stubbs, Harvard's Dean of Science from 2018 to 2024, has joined its Long-Term Benefit Trust. Stubbs helped discover the accelerating expansion of the universe, was the inaugural project scientist for what is now the Vera C. Rubin Observatory, and is a member of JASON, which advises the US government on defense and technology.
- Anthropic publishes its annual Public Benefit Report
Anthropic, a Public Benefit Corporation, published its annual Public Benefit Report, which sets out its five pillars and the actions it is taking to advance them.
- Anthropic says Claude designed 17 lab-confirmed protein binders against scorpion neurotoxins and releases all 211 tested designs
Anthropic says an internal research model working in Claude Science spent about 64 hours, mostly autonomously with four follow-up prompts, designing minibinders against scorpion neurotoxins, in a project developed with Citrini Research. Claude generated 318,484 designs and picked 180 for testing; Anthropic's team added 31 more, and Adaptyv Bio and Twist Bioscience lab-tested all 211. Anthropic says 17 designs met its criteria for binding one of four neurotoxins dangerous to humans, but it has not confirmed whether they neutralize venom in cells and does not claim an antivenom. All 211 designs and their data are being released, alongside a technical report.
- Anthropic launches API user profiles so platforms can tell it which end customer each request is for
Anthropic has launched user profiles in the Claude API, letting platforms, gateways and resellers tag which of their end customers each API request is for, so a safety enforcement action can apply to one end customer instead of the whole organization. The SDK and the ant CLI support the new /v1/user_profiles endpoints, and new organization webhooks in the Admin API (beta) report when a user profile is restricted or banned and when that ends.
- Dario Amodei responds to Nathan Calvin's critique that Anthropic puts racing ahead of safety
Anthropic CEO Dario Amodei replied on X to a long post by Encode's Nathan Calvin listing actions he said show Anthropic favoring its own interests. Amodei said he stands by Anthropic's "race to the top" strategy and its track record, and agreed Anthropic could have handled Calvin's items 7, 8 and 9 better, calling them communication that wasn't carefully written, and saying he takes responsibility. Among his counterexamples, he said Anthropic first had Mythos-class models in February but didn't make them generally available until June, when it had stronger cyber safeguards.
- Anthropic says it is hiring a smaller share of junior technical roles as agent hours outnumber employee hours 6 to 1
An Anthropic economics report using internal employment, AI-use, output, calendar and hiring data says the company has been hiring a smaller share of junior technical positions while hours worked by AI agents grew; by August, interactive agent hours outnumbered employees' 6 to 1. Anthropic says code per person, sales closed per rep and contracts processed all rose as agents spread, and that human attention became the bottleneck, with employees shifting to building visibility tools and monitoring agent pipelines. Hiring managers increasingly weigh autonomy and judgment over technical domain expertise, and Anthropic says it set up an early warning system for jobs at risk and prioritizes internal redeployment.
- Anthropic gives a six-person METR team employee-like access to evaluate its models, with METR free to publish its risk findings
Anthropic says it has onboarded six METR employees with company hardware, office space and access to internal tools roughly on par with its own risk assessment teams, excluding narrow categories such as customer data, privileged legal information and model architectures. METR will publish its analysis of Anthropic's risk levels, incidents and practices; Anthropic cannot veto publication but can require redactions in predefined categories, which METR may publicly characterize. Anthropic describes three tiers of evaluator access, with SecureBio in the first tier reviewing its August Risk Report's chemical and biological weapons chapter, Accenture and Faculty working toward a broader red-teaming partnership, and early talks with others including Hugging Face.
- Anthropic's Frontier Red Team says small AI models may be months away from self-replicating across cheap IoT devices
Anthropic researchers ran Gemma4-26B-A4B split across 15 off-brand Android TV boxes (about $55 each) at 0.39 tokens per second to show that a botnet of weak devices could host a capable model. In their evals, a 9B model got inference running on 8 of 9 hardware types, and GLM-5.2 and stronger models completed end-to-end replication in simplified tests. Models of 30B parameters and under can already do many individual replication steps but make mistakes chaining them. Anthropic estimates small models could reach this level by the end of the year if they keep improving at their historical rate, and closes with suggested defenses.
- Anthropic's new defenses cut distillation data harvesting by 90-99%, names Zhipu and Moonshot cases
Anthropic published new methods against distillation attacks, with early evidence they reduced the data harvested by distillers by around 90 to 99%. The post names Zhipu as having distilled the cyber capabilities of leading US models during GLM-5.3's development — after safeguards degraded its attacks on Claude Fable 5, it switched to an older Claude model and another lab's model — and a case where requests to Moonshot's Kimi models were secretly forwarded to Claude. CAISSI has assessed GLM-5.3 as the most cyber-capable open-weights model released to date, and Anthropic frames illicit distillation as both a safeguards gap and a national security risk through foreign military and surveillance integration.