What happened
On 22 October 2026, at AGNTCon + MCPCon North America, the Agentic AI Foundation (AAIF, the Linux Foundation-hosted neutral home for agentic-AI standards, currently governing MCP, A2A, AGENTS.md, goose and agentgateway) announced four new hosted projects, reaching 10 in its first year: Agent Skills (contributed by Anthropic — the open format that packages instructions, scripts and resources agents load on demand, with a proposed security-guidelines track), Open Responses (contributed by OpenAI — a multi-provider LLM interface specification), OpenEnv (contributed by Hugging Face — an interoperable RL environment framework), and OpenSandbox (contributed by Alibaba — a general-purpose sandbox platform with SDKs and a sandbox protocol for secure AI-agent execution, coding/GUI agents and RL training). The foundation said the projects cover 'where agents run safely' and are placed under open, vendor-neutral governance with multi-company maintainers.
Why it matters
AAIF is the neutral governance home for the standards that define the agentic-AI security-relevant attack surface and supply chain. Two of these projects are directly security-oriented: OpenSandbox establishes a shared sandbox platform/protocol for secure, isolated agent execution, and Agent Skills (packaging executable scripts/resources loaded by agents) ships alongside a proposed roadmap item for security guidelines for those who write and adopt skills. Consolidating MCP, Agent2Agent, AGENTS.md, skills and sandboxing under multi-stakeholder governance — with Anthropic, OpenAI, Hugging Face and Alibaba contributing and competitors as co-maintainers — shapes the baseline for agent identity, tool-permission and containment controls that practitioners will be expected to map into their agent-security frameworks.
Action needed
Track AAIF project governance and the emerging Agent Skills security-guidelines workstream; map MCP, A2A, AGENTS.md, skills and OpenSandbox projects into agentic-AI risk assessments and control catalogues (least-privilege tool access, sandbox isolation, supply-chain provenance for skills), and watch for a published security guideline/recommendation.