On 22 October 2026, Anthropic handed the keys to one of AI’s quietest success stories to the Linux Foundation. Agent Skills — the open SKILL.md standard that teaches AI agents specialist knowledge, now integrated into more than 70 platforms including ChatGPT, Cursor, GitHub Copilot and VS Code, with one million skills published and nearly 280 million installs — is now governed by the vendor-neutral Agentic AI Foundation (AAIF), which Anthropic co-founded with Block and OpenAI. The name, the logo, the GitHub organisation and the agentskills.io domain all moved. If you use an AI coding or work agent, this announcement affects you more than most this month — and it comes with a security warning worth reading first.
What did Anthropic actually announce?
Anthropic donated the Agent Skills standard — name, logo, GitHub organisation and the agentskills.io domain — to the Agentic AI Foundation under the Linux Foundation.
The announcement, published on 22 October 2026, frames the move as a commitment to keeping Agent Skills “open, community-driven, and vendor-neutral”. The Agentic AI Foundation (AAIF) is the same body Anthropic co-founded with Block and OpenAI last year, and to which it previously donated the Model Context Protocol (MCP) — the other standard quietly holding the AI-agent ecosystem together.
The details that matter:
- The transfer was announced at the AGNTCon + MCPCon North America conference in San Jose, alongside three other new projects: OpenAI’s Open Responses, Hugging Face’s OpenEnv and Alibaba’s OpenSandbox — taking the foundation to 10 hosted projects and 305 member organisations, according to WebProNews.
- Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. “The AAIF has done important work in giving the Model Context Protocol (MCP) a neutral home since we donated it less than a year ago,” Soria Parra said in the foundation’s release.
- You can keep using the standard or contributing to it without joining the AAIF — the donation changes governance, not access.
What are Agent Skills, in plain language?
A skill is a folder containing a SKILL.md file that teaches an AI agent how to do one specific task, plus any scripts or reference files it needs.
Anthropic says the idea came out of building Claude Code. The agent handled general knowledge work well but kept stumbling on the procedural knowledge real business tasks need — the step-by-step “how things are done here” that never makes it into training data. Skills are that missing instruction manual: a small, readable package an agent loads before it starts work.
The timeline is brisk. Skills first shipped in October 2025; within weeks, developers had published thousands on GitHub and written loaders for Codex, Cursor and OpenCode. Anthropic published the format as an open standard in December 2025, and the numbers since are striking: more than 70 agent platforms integrated, one million skills published, nearly 280 million installed (Anthropic, 22 October 2026). One popular framing — “write once, run on every big AI agent” — went viral in Hindi and Vietnamese tech explainers within a day of the donation news, a sign the concept is crossing from developer circles to mainstream users.
| Part of a skill | What it does |
|---|---|
| SKILL.md | The instruction manual — what the skill is for, when to use it, and how. |
| Scripts | Optional runnable helpers the agent can execute (this is the powerful — and risky — part). |
| References | Examples, documentation or data the agent consults while working. |
A quick clarification, because the names collide: Meta’s Muse assistant also ships built-in skill packs for small-business users — we covered Meta’s copy-paste skill packs here. Same SKILL.md format family, but a different product. This story is about the standard itself going vendor-neutral, not about a new feature in any one app.
Why does this handover matter to you?
Because the standard that already powers your agent’s add-ons now has a neutral home — vendor lock-in gets harder, and skills should get safer and more consistent to adopt.
If you use ChatGPT, Cursor, GitHub Copilot or VS Code, you benefit without doing anything: skills written for one agent increasingly work in the others, and a neutral foundation makes that interoperability harder for any single company to walk back. For developers, the payoff is on the roadmap: versioned releases (so a skill written today doesn’t silently break tomorrow) and a formal proposal process for specification changes, with discussion open to the community on the project’s Discord.
The broader pattern is worth noticing. MCP went to the same foundation less than a year ago; now Agent Skills follows. The two most important agent standards — one for connecting to live data, one for teaching procedures — now share a roof, maintained by engineers from rival labs. That is unusual in this industry, and it is good for users.
The catch: nearly a thousand skills carried malicious payloads
Skills ship plain-English instructions that an agent reads and acts on — which is exactly what makes a poisoned one dangerous.
JFrog’s 2026 Software Supply Chain Security State of the Union report, cited in the company’s 21–22 October announcements, tracked malicious agent skills and found 969 carrying high-impact payloads, alongside 495 malicious AI models on Hugging Face (JFrog, vendor-reported — treat the exact counts as the company’s findings, not independent measurement). The mechanism is simple and hard to unsee: a skill can hide instructions telling your agent to exfiltrate files or run commands, and your agent will obey because following the skill is its job.
“Skills and plugins ship plain-English instructions that an agent reads and acts on.”
— JFrog, October 2026 security release
This is why Anthropic’s proposed roadmap leads with security guidance for people who write and install skills. It is also why our standing advice applies double here: run through the seven checks before your agent goes rogue, and remember that banks are already blocking AI agents they don’t trust — the ecosystem is tightening from both ends.
Our take — the 3-question install checklist (desk-built, not hands-on tested): before installing any community skill, ask (1) Who published it? — a known organisation or well-starred repository beats a random fork; (2) What can it touch? — a read-only skill is a different risk class from one bundling scripts that run shell commands or reach your files; (3) Do you need it from the community at all? — built-in and verified catalogues first, random GitHub links last. Skills are the browser extensions of the agent era; install them with the same suspicion.
What happens next?
Versioned releases, a formal proposal process and security guidance lead the to-do list — and anyone can weigh in.
Anthropic has proposed starting the AAIF chapter with three workstreams: versioned releases of the specification, a formal proposal process for changes, and the security guidance described above. As with MCP, the project promises to prioritise community input and transparent decision-making, with proposed specification changes discussed on the project’s Discord. The full specification lives at agentskills.io.
What didn’t change: existing skills keep working, the format stays open, and no one needs to join anything to use it. Governance moved; your agents didn’t notice.
How is this different from MCP?
MCP connects agents to live data; Agent Skills teach them procedures — and both now live under the same Linux Foundation roof.
| Standard | What it standardises | Status |
|---|---|---|
| MCP (Model Context Protocol) | How agents connect to tools and live data sources | Donated to AAIF earlier; now at 10 hosted projects with Skills |
| Agent Skills | Reusable procedural knowledge in SKILL.md folders | Donated 22 Oct 2026; 1M skills, ~280M installs |
| Meta’s built-in skills | Product-specific packs for Muse small-business users | Separate product; same format family; unaffected by this move |
Frequently Asked Questions
Do I need to do anything after this donation?
No. If your agent platform already supports skills — ChatGPT, Cursor, GitHub Copilot, VS Code and 70+ others — everything keeps working. The donation changed who governs the standard, not how you use it.
Are community skills safe to install?
Treat them like browser extensions. JFrog’s 2026 research found 969 skills carrying high-impact malicious payloads (vendor-reported). Use the 3-question checklist above: check the publisher, check what the skill can touch, and prefer verified catalogues over random links.
What’s the difference between Agent Skills and MCP?
MCP standardises how agents connect to live tools and data; Agent Skills standardise reusable procedural knowledge — the “how to do the task” files. Anthropic donated both to the same Linux Foundation body, less than a year apart.
Does this change Meta’s Muse skills?
No direct effect. Meta’s built-in skill packs for Muse are a product feature in the same SKILL.md format family; this announcement concerns the open standard’s governance, not Meta’s product.
Where can I read the Agent Skills specification?
At agentskills.io, with discussion and contributions on the project’s GitHub organisation — both transferred to the Agentic AI Foundation as part of the donation.
Sources and methodology
Claims rest on Anthropic’s announcement “Donating Agent Skills to the Agentic AI Foundation” (published 22 October 2026, verified live 23 October 2026), secondary reporting on the AGNTCon launch and foundation membership, and JFrog’s 2026 Software Supply Chain Security State of the Union figures as cited in the company’s 21–22 October 2026 security releases (vendor-reported; counts are JFrog’s findings, not independently measured). Figures such as the 70+ platforms, one million skills and ~280 million installs are Anthropic’s stated numbers. Social virality was observed via read-only search of posts published 22–23 October 2026. Desk-researched; nothing here was hands-on tested in a live agent.
- Anthropic — Donating Agent Skills to the Agentic AI Foundation (22 Oct 2026)
- Agent Skills specification — agentskills.io
- WebProNews — Anthropic Hands Its Agent Skills Standard to the Linux Foundation’s Agentic AI Foundation (22 Oct 2026)
- UpTech Media — JFrog expands AI security tools to detect and govern AI assets (22 Oct 2026)
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.
Related posts
Is Muse AI Available in India? Status, Sightings & How to Join the Waitlist
Muse AI is not available in India as of October 2026 — only the US and Canada have it. Meta…
Read more
Walmart Welcomes AI Shopping Agents While Amazon Blocks Them
Two of the world's biggest retailers have just placed opposite bets on the same technology. On 21 October 2026, Walmart…
Read more
Gemini 4 Argon vs Carbon: What Is Actually Different?
Gemini 4 Argon and Carbon keep appearing in the same headlines, but they are not two competing products. One is…
Read more