The Agent Skills Directory

[Command Execution] (HIGH): The skill presents a high indirect prompt injection risk by directing the agent to run code. Evidence: (1) Ingestion: Processes user-provided Rust source and benchmark code as implied in patterns/optimization-guide.md. (2) Boundaries: No delimiters or instructions to ignore embedded content are provided. (3) Capabilities: The guide instructs the agent to use cargo bench, perf, valgrind, and heaptrack to execute or analyze code. (4) Sanitization: None provided. Running benchmarks on untrusted code results in arbitrary code execution by design.
[External Downloads] (MEDIUM): The skill recommends installing unverified third-party tools like flamegraph and cargo-instruments using cargo install, introducing unverifiable external dependencies into the environment.
[Prompt Injection] (LOW): The metadata description in SKILL.md uses the 'CRITICAL' marker to manipulate the agent's prioritization and tool selection logic, which is a form of instruction overriding.

m10-performance