The Agent Skills Directory

[SAFE]: The skill acts as a safety-oriented procedure for troubleshooting, providing the agent with strict boundaries to avoid common security pitfalls during failure recovery.
[CREDENTIALS_UNSAFE]: The skill incorporates a specific safety check in rules/record-the-learning.md that forbids the recording of secrets, tokens, or credentials in learning files, and mandates a search for credential-shaped patterns before saving records.
[PRIVILEGE_ESCALATION]: The rules/stop-conditions.md file explicitly instructs the agent to stop if a proposed fix requires widening permissions, weakening protections, or changing shared configurations, effectively preventing unauthorized privilege escalation.
[EXTERNAL_DOWNLOADS]: The skill provides defensive guidance in rules/stop-conditions.md against supply chain attacks by mandating that the agent stop if the owner of a tool or package is unknown, preventing installations from unidentified sources.
[INDIRECT_PROMPT_INJECTION]: Although the skill is designed to ingest and process potentially untrusted data from tool outputs and external documentation, it mitigates the risk of indirect prompt injection by enforcing a structured state machine (blocker/BLOCKED through blocker/CLEARED) and requiring direct verification of component-level claims.

treat-blockers-as-incidents