The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from browser-reported $csp_violation events. Properties such as $csp_blocked_url and $csp_original_policy are attacker-controlled and could contain malicious strings intended to influence the agent's logic.
Ingestion points: Data is read via execute-sql from the events table (specifically $csp_violation events).
Boundary markers: The instructions emphasize that the reported header is a lead and the code is the source of truth, providing a logical boundary for the agent.
Capability inventory: The skill uses execute-sql, gh (CLI), scout-emit-report, and scout-scratchpad-remember.
Sanitization: The skill performs normalization of nonces in CSP policies and explicitly instructs the agent to reconcile client-reported data with trusted source code from GitHub.
[COMMAND_EXECUTION]: The skill instructs the agent to use shell commands (gh, grep, base64) to fetch and analyze CSP policies from source repositories.
Evidence: Use of gh search code and gh api ... | base64 -d to retrieve repository contents. These operations are restricted to repositories identified via trusted, human-authored sources as per the skill's own logic.

signals-scout-csp-violations