The Agent Skills Directory

Indirect Prompt Injection (HIGH): The skill is designed to ingest and optimize external Rust code, creating a significant injection surface.
Ingestion points: The agent reads user-provided Rust source files, Cargo.toml configurations, and benchmarking data.
Boundary markers: No specific boundary markers or 'ignore embedded instructions' warnings are provided for the data being processed.
Capability inventory: The skill utilizes commands like cargo test, cargo bench, cargo build, samply, heaptrack, and valgrind. It also has the capability to modify source code.
Sanitization: There is no evidence of sanitization for user-provided project files. If a user provides a project with a malicious build.rs or Cargo.toml, the agent's suggestion to run cargo bench or cargo build would trigger arbitrary code execution on the local machine.
Command Execution (MEDIUM): The skill frequently suggests running system-level profiling and build tools (perf, samply, valgrind, cargo). While expected for a performance expert, these commands are powerful and execute with the user's local privileges.
Remote Code Execution (LOW): While no direct remote downloads are initiated by the skill itself, the reliance on cargo (which fetches dependencies from crates.io) means that malicious dependencies in a user's project could be downloaded and executed during the 'Measure First' or 'Benchmark' phases.
Metadata Analysis (SAFE): Metadata fields (name, description, license) are accurate and reflect the skill's actual behavior without deceptive instructions.

rust-performance