The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple local shell scripts and Python files (./bin/agent.sh, scripts/benchmarks/make_report.py, scripts/benchmarks/readme_snippet.py). It also performs Git operations including tagging and pushing to the remote repository.
[DATA_EXFILTRATION]: The skill explicitly accesses sensitive authentication files located at ~/.codex and ~/.claude. It also reads from and writes to the local filesystem (/tmp, evals/, benchmarks/), which is necessary for its stated purpose of benchmarking but represents a data exposure surface.
[PROMPT_INJECTION]: The skill contains instructional markers like 'Do Not Skip' and 'CRITICAL' (implicitly through 'Success Criteria') which attempt to enforce specific agent behavior during the benchmarking process.
[INDIRECT_PROMPT_INJECTION]: This skill presents an indirect prompt injection surface as it ingests and processes 'rows.jsonl' and journey data which likely contain untrusted or externally generated prompt/response text.
Ingestion points: Reads data from evals/journeys/*.json and /tmp/graphistry_skills_sweep_*/rows.jsonl.
Boundary markers: None explicitly defined in the provided file to separate processed data from instructions.
Capability inventory: Includes filesystem writes, network operations via Git, and execution of local shell and Python scripts.
Sanitization: Mentions a --public-safe flag for redacting source paths in reports, but does not specify sanitization of the input data itself to prevent injection.

benchmarks