The Agent Skills Directory

[PROMPT_INJECTION]: The skill identifies a potential attack surface for indirect prompt injection as it ingests and processes untrusted repository files during the recon and audit phases. This is addressed by 'Hard Rule 4', which instructs the agent to treat file contents as inert data and explicitly ignore steering attempts.\n- [PROMPT_INJECTION]: The agent is instructed to actively detect and report malicious patterns within analyzed files ('flag it as a finding and move on'), serving as a defensive layer against prompt injection attempts targeted at the model's context.\n- [COMMAND_EXECUTION]: The skill utilizes standard CLI utilities, including grep for codebase exploration and git rev-parse to capture the environment state (commit hash) for generated audit plans. These operations are limited to read-only analysis and documentation.\n- [PROMPT_INJECTION]: Mandatory evidence chain for Indirect Prompt Injection surface: 1) Ingestion points: The skill reads various codebase files (CSS, JS, TSX, etc.) using grep and file-read tools. 2) Boundary markers: Hard Rule 4 provides explicit instructions to ignore embedded instructions. 3) Capability inventory: The skill can write plan files to the /plans directory and spawn subagents for fanned-out audits. 4) Sanitization: The skill relies on instructional boundaries rather than programmatic sanitization.

improve-animations