The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Bash and Task tools to run external scripts and compiler utilities such as clang and cargo. These scripts perform tasks like extracting compile flags and emitting intermediate representations (IR) or assembly for analysis.
[DYNAMIC_EXECUTION]: As part of its verification phase, the skill generates proof-of-concept (PoC) source files and test harnesses. It then compiles and executes these programs on the local system to validate whether a detected zeroization issue is exploitable.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted source code and build metadata, which represents an indirect prompt injection attack surface. 1. Ingestion points: The skill reads local C/C++/Rust source files and build configuration files (e.g., compile_commands.json, Cargo.toml) provided by the user. 2. Boundary markers: While the instructions explicitly tell the agent to ignore deceptive arguments found in code comments, there are no defined structural delimiters for the untrusted code content itself. 3. Capability inventory: The agent possesses the capability to execute shell commands and compile/run generated code via the Bash and Task tools. 4. Sanitization: The skill does not specify any automated sanitization or filtering of the audited code before it enters the agent's context.

zeroize-audit