Agent Tool Risk

Use this skill for MCP, plugins, hooks, sub-agents, tool permissions, prompt flows, automation loops, and model-visible logs.

Workflow

1. Inventory the agent/tool surface: trigger, command, permissions, inputs, outputs, and persistence.
2. Classify risk:

- prompt injection or untrusted content - excessive agency or destructive authority - secret exposure - data exfiltration through logs or MCP - command injection - stale or misleading model-visible memory

1. Confirm high-risk operations have human approval or guard hooks.
2. Ensure hidden logs are not loaded as default model context.
3. Ensure model-visible logs contain only durable, necessary retry context.
4. Update docs/harness/AGENTSECURITY.md, docs/harness/SECURITYPOLICY.md, and docs/harness/SUBAGENT_PROTOCOL.md if behavior changes.

Required Output

Return tool surface, risk class, mitigations, required approvals, logging impact, and remaining risks.

Blockers

• Secrets written to repo, hidden logs, or model-visible memory.
• Mutating cloud or git actions without explicit approval.
• Hooks that execute untrusted user content as shell commands.
• MCP credentials committed or printed.