The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze untrusted source code from various projects, which creates an inherent attack surface for indirect prompt injection. Malicious instructions could be hidden in the scanned code (e.g., inside comments or string literals) to influence the agent's behavior or misrepresent findings. * Ingestion points: The skill reads project files throughout its workflow, specifically during Step 1 (Scope Resolution), Step 3 (Secrets Scan), and Step 4 (Vulnerability Deep Scan) as defined in SKILL.md. * Boundary markers: Step 6 (Self-Verification Pass) in SKILL.md mandates a reasoning step where the agent re-evaluates its findings to filter false positives, which serves as a logical guard against injection attempts. * Capability inventory: The agent reads files and is instructed to run local dependency audit tools like npm audit or pip-audit. * Sanitization: There is no mention of automated sanitization or escaping of the input code content; the skill relies on the agent's reasoning capabilities.
[SAFE]: Static analysis triggers for eval() and exec() in references/language-patterns.md and references/vuln-categories.md are benign false positives. These references provide educational examples of what the agent should look for when auditing other codebases and do not represent executable code within the skill itself.

security-audit