🧭 Anthropic Donates Agent Skills to the Agentic AI Foundation — 70+ Platforms, 280M Installs

Anthropic has donated Agent Skills — the open standard for giving AI agents specialised procedural knowledge — to the newly formed Agentic AI Foundation (AAIF), a neutral, Linux Foundation-hosted organisation co-founded by Anthropic, Block, and OpenAI. The move transfers governance of the standard from a single vendor to a multi-stakeholder body, ensuring that the 70+ agent platforms already integrated — including ChatGPT, Cursor, GitHub Copilot, and VS Code — can participate in the standard's future without lock-in to any one AI provider.

What Agent Skills is

An Agent Skill is a folder containing a SKILL.md file that teaches an agent a specific task or domain, plus any supporting materials (templates, example files, configuration stubs). The format originated in Claude Code, which needed a way to package procedural knowledge that goes beyond what a model's training can reliably provide. Since launch in December 2025, the ecosystem has grown to 1 million published skills and nearly 280 million installs — a pace that, according to AAIF founding documents, already rivals the npm package count at the same stage of its lifecycle.

Governance structure

• Lead Maintainers: David Soria Parra (Anthropic) and Mahesh Murag (Anthropic) hold the initial roles; additional maintainer slots will be opened to the community via a transparent nomination process.
• Proposal process: Changes to the standard go through a formal RFC-style process with public comment periods, mirroring how web standards and language specifications evolve.
• Security guidance: The AAIF roadmap includes formalised security guidance for skill authors — particularly around skills that invoke shell commands or perform file writes — which has been a gap in the early ecosystem.
• Versioning: The standard will move to explicit version identifiers, allowing platforms to declare compatibility and agents to resolve the appropriate skill version for their runtime.

🧭 Claude Code v2.1.307-309: Sandbox Security Hardened and Safety OTel Event Lands

Three Claude Code releases shipped between October 21 and 22 — v2.1.307, v2.1.308, and v2.1.309 — delivering the most substantial sandbox security overhaul since the Linux sandbox launch. The headline addition in v2.1.307 is a new claude_code.safety_event OpenTelemetry event emitted for every tool call, giving teams observability into when and why Claude invokes tools without having to parse log text. v2.1.309 then closed a cluster of sandbox escape vectors that had been reported by the security community.

Sandbox security fixes in v2.1.309

• Shell startup files: Sandboxed Bash commands on macOS and Linux were previously running shell startup files (.bashrc, .zshrc) outside the sandbox boundary. This meant any alias or function defined there could access the host filesystem. Fixed.
• Symlink bypass: macOS sandbox denyWrite glob patterns were not resolving symlinks, allowing a write to a symlinked path to bypass the deny rule. Fixed.
• MCP config protection: Sandboxed commands could previously modify MCP config files, which would persist after the session. MCP config paths are now explicitly write-protected within the sandbox.
• Compound command permissions: Bash permission checks were not correctly applying to compound commands that contained directory changes (e.g. cd /tmp && rm -rf ...). The permission check now applies to the full resolved execution context.
• Worktree git isolation: git replace refs were leaking across worktree boundaries. Isolated.

v2.1.307 developer additions

• /claude-api api-version command: Moves the Claude API code in your workspace to use API version 2026-10-21. Run this to adopt any new API features introduced in the October version.
• claude_code.safety_event OTel event: Emitted on every tool call, carrying the tool name, whether it was approved or denied, and the reason. Integrate with your existing OpenTelemetry collector to build dashboards for tool-use patterns across your team without any log parsing.
• Fixed autocompact thrashing: Sessions with large tool results were triggering repeated compaction cycles. Resolved by tracking compaction state across the context-length check.

🧭 Anthropic Opens Life Sciences Guest Investigator Program for Academic Researchers

Anthropic has posted an open call for guest investigators in the life sciences — an 8–12 week collaboration programme in which principal investigators and their research groups at academic institutions, nonprofits, or government labs work directly alongside Anthropic researchers using Claude for scientific discovery. Anthropic covers all computational costs; participants retain full data ownership and IP rights to any independent follow-up work. All results must be published and open-sourced.

Research focus areas

• Causal human biology (identifying mechanistic relationships, not just correlations)
• Structural and computational molecular biology
• Medicinal chemistry and drug discovery
• Clinical science (evidence synthesis, trial design support)

Why this matters beyond the life sciences

The programme is the first time Anthropic has offered embedded collaboration — teams working inside the company for a fixed term — rather than arm's-length API access. The deliverable is not a paper but "analyses, artifacts, and tools" that demonstrate preliminary evidence within the collaboration window. The constraint on open-sourcing results is notable: Anthropic gets early exposure to high-value scientific workflows, researchers get frontier compute and model access, and the broader community gets open artefacts. It is structurally similar to how national labs run visiting researcher programmes, applied to AI-augmented scientific work.