The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze untrusted source code and architecture documentation. These inputs can contain malicious instructions hidden in comments, documentation, or string literals intended to influence the agent's behavior or bias the audit results. The skill does not currently include explicit boundary markers or instructions for the agent to ignore commands found within the analyzed data.
Ingestion points: Full source code repository access and architecture documentation specified in SKILL.md.
Boundary markers: None present. The instructions do not explicitly tell the agent to treat the analyzed code solely as data.
Capability inventory: The skill uses tools defined in references/tools.md (semgrep, CodeQL, ripgrep, jadx, sourcemap-extract) which involve subprocess execution.
Sanitization: No validation or filtering is performed on the content of the analyzed source code files.
[EXTERNAL_DOWNLOADS]: The sourcemap_extract tool defined in references/tools.md is designed to fetch content from external URLs. While this is a legitimate technique for recovering original source code from minified JavaScript, it involves network interaction with potentially untrusted external servers based on provided URLs.

pentest-whitebox-code-review