KDCube
← AI Industry News

Agent Skills Goes Neutral, Logins Go Passwordless

Anthropic handed the Agent Skills standard to the vendor-neutral Agentic AI Foundation, now co-maintained with Google and OpenAI, while 1Password for Agents shipped passwordless, user-approved logins for agents. A METR + Redwood red-team also found OpenAI's internal monitors could be bypassed to hide h...

Highlights

top scan
  • Anthropic donated the Agent Skills standard — name, logo, GitHub org, and the agentskills.io domain — to the Agentic AI Foundation under the Linux Foundation, to be jointly maintained by engineers from Anthropic, Google, and OpenAI. Reported adoption: 70+ platforms, ~1M skills published, nearly 280M installs since December 2025. (Anthropic)
  • 1Password for Agents shipped: agents log in on your behalf without ever seeing credentials, using a per-agent signed identity plus an explicit user approval step. Launch partners span Claude Duet, ChatGPT Work, Grok, Meta's Muse, and Town. (TechCrunch)
  • A METR + Redwood Research red-team embedded in OpenAI found exploits that could hide hundreds of agent actions from internal safety monitors — and estimated only ~75–94% of GPT-5.5-class traffic was even eligible for monitoring. (CNBC)
  • vLLM added day-0 support for MiniMax-M3.1-Flash on Oct 23, keeping the open-serving lane moving even as the week's headlines skewed toward governance. (vLLM)

Key Signals

ranked scan

  1. Agent Skills becomes vendor-neutral infrastructure

    Oct 22 · Anthropic

    With Skills now at the AAIF alongside MCP, two of the core agent primitives — tools and reusable skill packages — sit under the same neutral foundation with cross-vendor lead maintainers. For builders, that lowers the risk of betting on a format that a single vendor could later fork or gate. (Anthropic, Simon Willison)

  2. Your agent monitoring probably has blind spots

    Oct 22 · METR/Redwood via CNBC & Reuters

    The disclosure matters less for OpenAI specifically (which says the production gaps are patched) than for the general lesson: coverage of agent actions is partial by default, and a determined agent path can route around it. METR notes it has not verified the fixes. (CNBC, Reuters)

  3. Credential brokering is the new agent-identity front

    Oct 22 · 1Password

    Where this week's prior issue tracked enterprise agent passports (ServiceNow, watsonx), 1Password pushes a consumer-grade broker: the agent proves identity cryptographically, the human approves, and the password never leaves the vault. Same trust problem, a different layer of the stack. (TechCrunch)

Market Heat

topic map
Skillsstandard
MCPtools
Govmonitoring
Authidentity
Agentsruntime
LLMserving
Opsaudit
Trustcontrols

Why It Matters / What To Watch

operator actions
  1. Standardize on the primitives that just went neutral.
    • If you maintain Skills or MCP servers, track AAIF governance and the new joint-maintainer cadence rather than Anthropic's repo alone. (Anthropic)
    • Watch whether OpenAI and Google ship first-class Skills tooling now that they co-maintain it — portability across ChatGPT, Copilot, and Claude is the payoff. (Simon Willison)
  2. Audit what your own runtime can actually see.
    • Don't assume 100% of agent actions are logged or inspectable; measure eligible-vs-monitored coverage for your highest-capability models. (CNBC)
    • Treat "monitor can be bypassed" as a threat model, not an edge case — pair observability with hard runtime controls (approvals, scoped credentials). (Reuters)
  3. Decide where agent credentials should live.
    • Evaluate broker-style auth (user-approved, per-agent signed identity, no credential exposure) before letting agents hold raw secrets. (TechCrunch)

Quick Links

sources