The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts using python3 .plan/execute-script.py to interact with project management tools. These operations include resolving worktree paths, computing file footprints, and detecting architectural modules. All scripts are internal to the .plan directory of the repository and do not involve remote code downloads or untrusted network operations.
[DATA_EXPOSURE_AND_EXFILTRATION]: While the skill accesses file paths to perform a security audit, it does not exfiltrate this data. Findings are emitted locally into the manage-findings tool, and the execution is confined to the plan-bound environment. No external network requests or hardcoded credentials were detected.
[PROMPT_INJECTION]: The skill includes instructions to guide the LLM through a 'cognitive audit' in Stage 4. These instructions are focused on security principles (OWASP, STRIDE) and include explicit constraints to prevent the agent from deviating from the five-stage procedure or creating unauthorized finding types. There are no attempts to bypass safety filters or override system-level constraints.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because Stage 4 involves an LLM review of 'in-radius files' (the footprint). If these files contain malicious instructions, they could influence the audit's cognitive process.
Ingestion points: Stage 4 reads in-footprint files derived from the git diff (Stage 1).
Boundary markers: The skill lacks explicit 'ignore embedded instructions' markers for the content being audited, though it relies on the persona-security-expert context to guide analysis.
Capability inventory: The skill has capabilities to write findings via manage-findings add and resolve paths via manage-architecture.
Sanitization: No specific sanitization of the audited file content is mentioned before it is processed by the LLM in the cognitive audit stage. Severity is assessed as LOW due to the nature of this category.

recipe-security-audit