Agent SkillsSupply ChainReadiness Signal
Anthropic's official document skills load a file from a shared folder without checking who put it there
Air Security reports that the helper behind the Word, PowerPoint and Excel skills preloads a library from a fixed temp path; Anthropic is reported to consider it outside its security boundary.
Talking point
This is not a clever attack on a model. It is an old pattern: a trusted helper runs whatever file it finds in a shared temp folder. Air Security says the document skills load a library from a fixed path with no check on who wrote it, so anything else that can write to that folder could run code as the user. The vendor reportedly sees it as outside its boundary because it only happens inside a sandbox, and both positions are worth knowing. What generalizes is the lesson: a skill is code your agent runs, and 'official' tells you who wrote it, not that nobody should review it.
Content angle
Post: 'A skill is software.' Frame it for non-technical leaders: you would not let an employee install an unreviewed program from a shared drive, yet agent skills, plugins and extensions are exactly that. Ten-minute workshop: list every skill, plugin and extension installed on one team's agents, who approved each, and who owns updates. Pair it with last week's coding-agent configuration card and the human-layer point that people install whatever makes the work easy. DoCRA-flavored angle (my framing, not new DoCRA writing): reasonable care includes knowing what code your agents execute.
Source: Air Security: Office Party, a flaw in Anthropic's official skills; Air Security: OpenAI Sesame (October 14, 2026)
NISTAgent IdentityAccess ControlReadiness Signal
NIST proposes ten design tenets for agent identity, and says guardrails alone are not enough
An NCCoE blog post offers an initial checklist: distinct agent identities, short-lived access, policy separated from the model's reasoning, deterministic access decisions, and tamper-evident logs the agent cannot edit.
Talking point
NIST's agent identity team says what many security leaders have been circling: because a model takes instructions and data in one stream, the line between what an agent tries to do and what it is allowed to do has to be built into the architecture, and 'guardrails alone are insufficient.' Two of the ten tenets are the ones to remember. Access decisions should be deterministic, with model judgment only as an input, and the record of what an agent did has to live where the agent cannot edit it. It is a proposal, not a standard, but it is a ready-made list for asking a vendor what they actually built.
Content angle
Post or talk segment: 'The agent is not the policy.' Pick tenets seven, eight and ten and turn each into one question a CEO can ask a vendor. Use the contrast between a prompt that says 'do not' and a permission the agent simply does not have. Workshop: take one agent in the business and answer all ten tenets as yes, no or don't know; the 'don't know' count is the finding. Ties to Pilot Fixer, since pilots stall when nobody owns the permission layer.
Source: NIST: Design Tenets for Agent Identity Management Systems