agent-securitylisted

# Agent Security — audit skill Generic skill shipped with the training. Walks two lenses against any agent system pointed at it: 1. **Access-control analysis** — what external systems the agent can reach, and whether each reach is necessary. 2. **Agent-STRIDE** — the six STRIDE categories adapted to agent systems (not classical software systems). ## Invocation Ask Claude to "apply the agent-security skill to [target system]." Output lands at `module-4/security-report.md`. ## Access-control pass For each external touch-point the target system has, answer: - What does the agent reach? (connector, file write beyond working dir, sub-agent dispatch, tool use) - Why does it reach? (what the system actually does with it) - Could the system work without it? (if yes → unused access, flag) List unused access as a cluster of access-control findings. Each gets a severity rating (high / medium / low) based on what the access could enable if the agent misbehaves. ## Agent-STRIDE Six categories, each adapted. See `stride.md` for full definitions. Short form: - **Spoofing** — an agent impersonates a human author; a memory file pretends to be from a source that doesn't exist. - **Tampering** — an agent modifies memory / rules / sources in ways that sneak into future sessions; output drift through silent edits. - **Repudiation** — the agent writes without attribution; no audit of which agent produced what at what time. - **Information disclosure** — the agent reveals c