Anthropic donates Agent Skills to the Agentic AI Foundation, details its distillation defenses, and puts 5% of its shares into a public benefit fund

Anthropic is giving the Agent Skills open standard — its name and logo, the GitHub organization and the agentskills.io domain — to the Agentic AI Foundation (AAIF), the Linux Foundation body Anthropic co-founded with Block and OpenAI and where it already donated the Model Context Protocol. Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated skills, with about 1 million skills published and nearly 280 million installs. Engineers from Anthropic, Google, OpenAI and others will maintain the spec as individuals, led by David Soria Parra and Mahesh Murag.

Why it matters now: skills are the portable unit of agent know-how, and that portability only holds if no single vendor can change the format. Moving it next to MCP completes a two-standard stack (tools and procedures) under one neutral roof. The proposed roadmap — versioned releases, a formal proposal process and security guidance for people who write and install skills — is the part to watch, since a million-skill ecosystem is also a supply-chain surface.

Anthropic lists six layers: summarized reasoning, preserved thinking (which stops new API accounts editing Claude’s prior context, introduced with Fable 5.1 and Opus 5.5), reasoning isolation so transcripts are readable only by the organization that produced them, a two-stage classifier, stricter limits and bans for repeat extraction attempts, and charging for requests blocked before Claude responds. It reports extracted reasoning tokens down 10–100x+ on its platform, a false-positive rate under 0.1%, and tens of thousands of distillation-linked accounts banned. In a test, a model trained on full Opus 4.8 transcripts scored 80.4% on SWE-bench Verified versus 71.8% on summarized ones.

Why it matters now: this is the clearest public number yet on what reasoning traces are worth — roughly eight benchmark points — which explains why labs now hide them. Anthropic also names Zhipu (GLM-5.3) and Moonshot (Kimi) as cases, and calls the effort a cat-and-mouse game. For builders, the practical edge is that preserved thinking required some integrations to change.

The Anthropic Public Benefit Fund is a wholly owned LLC, structured so it can make large donations to nonprofits or direct distributions to specific populations. Anthropic commits 5% of its shares, with a possible further 2% at the board’s discretion for every 50% increase in company value above a baseline; with earlier commitments, roughly 20% of shares are earmarked for the public good. None of the Fund’s equity will be used for lobbying or political giving. Governance and valuation details are promised “in the coming months.”

Why it matters now: it is a bet that the labs’ main social contract will be paid in equity rather than taxes or licences, with AI labor-market disruption named as the first target (Anthropic cites an earlier $200 million commitment). Until the governance terms land, the open question is who decides how the stock is deployed.

Principal investigators at academic, nonprofit and government labs can apply to run 8–12 week projects with Anthropic’s research team, focused on causal human biology, molecular biology, medicinal chemistry, drug discovery and clinical science. Anthropic runs its models and harnesses at no cost, outputs are published with data and code open-sourced, and guests keep rights to their data and independent follow-on work. The window is about 14 days from Oct 22, so it likely closes in early November; the page gives no exact date.

Why it matters now: it turns the “AI for science” credits programme into hands-on collaboration, and the open-output rule makes it a source of public benchmarks for agentic research workflows. Apply soon if relevant.