The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted source code from external repositories. This creates a vulnerability surface where maliciously crafted comments or code patterns could attempt to influence the logic of the AI agent roles performing the audit.
Ingestion points: All source files and build configurations (compile_commands.json, Cargo.toml) within the target repository.
Boundary markers: The skill uses structured JSON schemas and markdown report templates to separate data from execution instructions, though it relies on LLM reasoning for much of the analysis.
Capability inventory: The skill possesses extensive capabilities including file system read/write, shell command execution (Clang, Cargo, Python), and role delegation.
Sanitization: Scripts like tools/extract_compile_flags.py and tools/track_dataflow.sh use standard quoting and escaping when processing strings for shell or JSON output.
[DYNAMIC_EXECUTION]: To validate its findings, the skill generates bespoke proof-of-concept (PoC) programs in C and Rust. These programs incorporate function signatures, variable names, and logic derived from the audited codebase. The skill subsequently compiles and executes these PoCs in the local environment to confirm whether sensitive data persists in memory.
[EXTERNAL_DOWNLOADS]: The C/C++ semantic analysis phase involves fetching and running the Serena MCP server from a public GitHub repository (github.com/oraios/serena) via the uvx package runner. This is a standard dependency for the tool's advanced analysis capabilities.

zeroize-audit