The Agent Skills Directory

[DYNAMIC_EXECUTION]: The benchmarking helper library provided in SKILL.md employs the eval command within core functions such as bench_time_ms and bench_warmup. This is used to execute the shell commands provided for timing and warmup iterations. This pattern facilitates arbitrary command execution within the generated scripts.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of scripts that benchmark shell tools by executing command strings. This creates an attack surface where untrusted data, if incorporated into the command string by the agent, could lead to command injection due to the lack of sanitization before the eval call in the benchmarking helpers.
Ingestion points: Command arguments passed to the bench_run, bench_warmup, and bench_time_ms functions in SKILL.md.
Boundary markers: The templates do not include boundary markers or warnings to isolate the command execution context from potentially untrusted data.
Capability inventory: The generated scripts have broad capabilities including shell command execution, file system modification (mkdir, rm, mktemp), and environment variable manipulation.
Sanitization: There is no evidence of sanitization or input validation for the command strings before they are executed.

performance-benchmark-specialist