The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted source code from external repositories, which serves as a potential vector for indirect prompt injection. Maliciously crafted code, comments, or documentation could attempt to influence the agent's logic or report generation during the multi-phase analysis process.
Ingestion points: The 2-source-analyzer and 2b-rust-source-analyzer agents read source files and documentation directly into the agent context for Phase 1 analysis.
Boundary markers: The instructions do not explicitly define delimiters or specific 'ignore' instructions to wrap the ingested source code data.
Capability inventory: The agent has extensive capabilities including file writing (Write), shell command execution (Bash, Task), and the ability to compile and run generated code (5b-poc-validator).
Sanitization: No specific sanitization, escaping, or filtering of the ingested source code is described before it is processed by subsequent agents.
[DYNAMIC_EXECUTION]: The skill pipeline includes phases specifically designed to generate, compile, and execute Proof-of-Concept (PoC) programs based on analysis findings.
Evidence: The 5-poc-generator agent crafts bespoke programs to demonstrate vulnerabilities, and the 5b-poc-validator agent compiles and runs these programs to verify findings.
Mitigation: The skill includes a 'Mandatory confirmation gate' that requires the agent to disclose the exact commands and wait for explicit user approval before performing any exploitative or probing actions.
[COMMAND_EXECUTION]: The analysis workflow relies on executing various system tools and internal scripts via the Task and Bash tools.
Evidence: The skill executes compiler tools like clang and cargo, as well as internal scripts such as emit_ir.sh, emit_asm.sh, and semantic_audit.py to perform deep analysis of the target codebase.

zeroize-audit