The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to orchestrate its 8-phase analysis pipeline. It executes numerous internal scripts (e.g., emit_ir.sh, emit_asm.sh, validate_rust_toolchain.sh) and external toolchains like clang, cargo, and rustc to perform its analysis.
[REMOTE_CODE_EXECUTION]: In Phases 4 and 5, the skill utilizes agents 5-poc-generator and 5b-poc-validator to create and subsequently execute Proof-of-Concept programs on the host. This runtime execution of generated code is a significant capability that could be exploited if the generation logic is influenced by malicious input code.
[EXTERNAL_DOWNLOADS]: The skill documentation references the use of uv and uvx for running Python-based analysis scripts (semantic_audit.py, find_dangerous_apis.py, etc.). These tools may fetch and manage external dependencies from public registries during the analysis process.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted source code from external repositories. This data is then used to influence the generation of PoCs and the final analysis report. The skill lacks explicit boundary markers or sanitization steps for the ingested data before it influences high-privilege operations like code execution.

zeroize-audit