What Anthropic handed to the Agentic AI Foundation
Anthropic has donated Agent Skills, the open format for packaging instructions and scripts that an AI agent can load, to the Agentic AI Foundation (AAIF), a body under the Linux Foundation that Anthropic says it co-founded with Block and OpenAI. The donation post published on 22 October 2026 covers three things: the Agent Skills name and logo, the GitHub organisation, and the agentskills.io domain, where the specification lives. The post describes no transfer of copyright or code, and the specification repository is licensed under Apache-2.0 for its code and CC-BY-4.0 for its documentation. Anthropic says anyone can keep using or contributing to the standard without joining the AAIF.
Agent Skills is a format for describing instructions to a model. MCP, the protocol that connects applications to external systems, is a separate AAIF project and a different kind of thing.
The Linux Foundation announced four new AAIF projects the same day: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. That takes the foundation from three projects to ten in its first year. “Today we host 10,” said Manik Surtani, the AAIF’s CTO and co-founder. The news came at AGNTCon + MCPCon North America in San Jose on 22 and 23 October.
One loose end sits on the foundation’s own pages. aaif.io/projects/skills links its GitHub icon to github.com/anthropics/skills, the example-skills repository, while the specification repository is github.com/agentskills/agentskills. That may be a stale link rather than evidence about the transfer, but it is the kind of detail a handover of this size leaves behind.
Whose numbers these are: Anthropic’s 70+, 1 million and nearly 280 million
Anthropic says more than 70 agent platforms have integrated Agent Skills, including ChatGPT, Cursor, GitHub Copilot and VS Code; that 1 million skills have been published; and that nearly 280 million have been installed. All three are Anthropic’s own figures, published without a methodology, and no independent count exists to check them against.
Who governs the standard, and who maintains it
By October 2026 the AAIF had published a governing board of ten named people, a technical committee of eight, priced membership tiers, ten projects and eight standing working groups. The board is chaired by David Nalley of AWS and includes Anthropic’s IP counsel, Linzy McCartney. The technical committee is chaired by David Soria Parra, an Anthropic staffer and a lead maintainer of MCP. Membership costs money: Platinum at $350,000 carries a guaranteed board seat, Gold at $200,000 brings nomination rights for a shared seat, Silver runs from $5,000 to $95,000 by headcount, and approved non-profits, academic and government bodies can join as Associates without voting rights.
Anthropic says engineers from Anthropic, Google, OpenAI and other companies will maintain Agent Skills as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. MCP’s governance, the stated model for this arrangement, is explicit that technical membership attaches to a person rather than an employer and that no seats are reserved for specific companies.
Two documented facts pull in different directions. Technical governance runs on individuals. Funding governance runs on companies: the board seat that comes with Platinum membership belongs to an organisation, and the committee that oversees the standard is chaired by someone employed by the company that wrote it.
On ownership, Anthropic’s post is the only source stating that the Agent Skills name and logo changed hands. AAIF’s brand page asserts that AAIF owns AAIF’s trademarks, which says nothing by itself about the Agent Skills mark.
What a skill contains, and what a client loads
A skill is a folder. It must contain a SKILL.md file, and it may also contain a scripts/ directory with executable code, a references/ directory with documentation and an assets/ directory with other files. SKILL.md opens with YAML frontmatter: a required name of up to 64 characters in lowercase letters, numbers and hyphens that must match the folder name; a required description of up to 1024 characters; and optional license, compatibility, metadata and allowed-tools fields.
Skills load in stages, which the specification calls progressive disclosure. At startup the agent reads only the name and description, roughly 100 tokens. When the agent decides a skill applies, it loads the full SKILL.md body, which the specification recommends keeping under 500 lines. Scripts and reference files load while the skill runs. The client-side code that reads a skill folder and offers it to a model is what Anthropic calls a loader; the specification site calls the same thing a client implementor.
The field that looks most like a permission control is the least settled. allowed-tools is a space-separated string of tools a skill asks to have pre-approved, and the specification marks it Experimental, warning that support may vary between agent implementations.
What the specification checks today
The specification’s validation section names one tool: the skills-ref reference library, run as skills-ref validate ./my-skill. It checks that SKILL.md frontmatter is valid and follows the naming conventions, so the check covers format. The library’s README is blunter, describing it as “intended for demonstration purposes only” and “not meant to be used in production”.
The specification defines no trust, permission, signing, provenance or install-time review model. There is no certification scheme, no signing key for a reader to verify against, and no reviewer standing between a published skill and an agent that runs its scripts. The Client Showcase lists agent products that support the format and makes no review, curation or certification claim.
The published rules for loading a skill safely sit in MCP
Skills served over the Model Context Protocol are covered by a different document. MCP’s Skills extension, SEP-2640, reached Final status on 13 September 2026, and its published overview sets out requirements the Agent Skills format does not carry.
Under Integrity and Verification, a server publishes SHA-256 digests and byte sizes for a skill’s files in a manifest. Approval binds to the full set of URIs and digests, and adding, changing or removing any file revokes that approval. Caches must be immutable or verified on every access, and cached files stay out of filesystem skill-discovery paths. Under Security Considerations, hosts must prevent a skill from silently replacing another of the same name, tag loaded content with the server it came from, require explicit per-call approval for reads across servers, and treat skill content as untrusted input: host-side code execution and permission grants such as allowed-tools require explicit approval from the user for each skill. Fresh consent is required before a nested skill activates.
The extension is careful about what a digest proves: “Digests establish consistency with the server’s manifest, not trust in its content.” It applies only to skills served over MCP. A folder copied from GitHub or a third-party marketplace sits outside these rules, and the overview notes that SDK and host support is still being implemented.
The foundation’s own skills security work is a draft
The AAIF does have a Security & Privacy working group: open, chaired by Alexander Frazer of Runlayer, with Junjie Bu of Google as vice chair and 224 active members. Its deliverables folder lists two documents, both marked Draft and both last modified on 30 September 2026.
The draft best-practices guide, marked “Draft v0.1, for WG review”, proposes a four-tier model for verifying skills. The first tier is an automated pre-scan that disqualifies anything with a critical or high finding. The second is a manual structural review of skills that request elevated permissions, checking allowed-tools against the stated purpose and scanning bundled scripts for undeclared domains, obfuscated code and paths reaching toward ~/.ssh or ~/.aws. The third is an organisational trust registry that pins versions and hashes to a named owner and re-verifies on update. The fourth is runtime monitoring against the scope the skill declared in allowed-tools.
The draft also names a consent problem: approving a skill’s first execution can authorise a wider scope of action than the preview showed, and that approval persists as a standing grant until someone revokes it. As a working document, none of this is AAIF policy yet, and it may change before publication. No AAIF working group is dedicated to skills. The Identity & Trust group, with 270 active members, works on portable agent identity and delegation, and publishes no skills document.
What published research finds in distributed skills
Snyk published a study called ToxicSkills on 5 February 2026. It scanned 3,984 skills from ClawHub and skills.sh, found at least one critical issue in 13.4% of them and some security issue in 36.82%, and confirmed 76 malicious payloads after human review, with eight malicious skills still live at publication. Snyk markets commercial AI-security products in the same post alongside the free scanner it recommends, so its figures come from a vendor with a commercial interest in the problem.
Two preprints add separate measurements, and neither has been peer reviewed. One, submitted on 15 January 2026, analysed 31,132 skills from two marketplaces and found at least one vulnerability in 26.1%, with 5.2% showing high-severity patterns that suggest malicious intent. A June 2026 preprint measured skill-injection attacks against terminal-based agents, reporting 81.4% success without a guardian and 18.6% with a dynamic guardian.
What “proposed” covers in the roadmap
Anthropic’s post lists three items it has proposed for Agent Skills under the AAIF: versioned releases, a formal proposal process, and security guidance for people who write and install skills. Community input runs through the project’s Discord. Proposed means exactly that: the standard’s new owner has not published any of the three, so the guidance that would tell an author what to check before publishing and an installer what to check before running a skill is a plan rather than a document.
What an installer can rely on today
As the documents stand:
- The Agent Skills specification validates frontmatter and naming through skills-ref, a reference library its own authors describe as unsuitable for production use.
- No signing, review, certification or permission standard applies to skills distributed as folders or through third-party marketplaces such as ClawHub and skills.sh. Neither marketplace is an AAIF project or run by Anthropic.
- Skills served over MCP carry normative rules on digests, origin, cache verification and per-skill approval. Those rules establish consistency with the server’s manifest, and they say nothing about whether the content is safe.
- Snyk, which sells AI-security products, reports a live supply-chain problem on the marketplaces it scanned and recommends scanning installed skills with mcp-scan, rotating credentials and inspecting memory files.
- The AAIF Security & Privacy working group’s draft proposes version and hash pinning with a named owner and re-verification on update. It is not adopted.
- The AAIF roadmap item that would give authors and installers agreed security guidance is still proposed.
