The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script evals/files/lazy/scripts/fetch-secret.sh contains a hardcoded credential-like string sk-live-5f1c9e0b7a. Although this appears to be a mock value used to verify that the skill can correctly handle and protect secrets during deployment tasks, hardcoding strings that match sensitive patterns can trigger security alerts.
[DYNAMIC_EXECUTION]: The test utility evals/files/testargs/scripts/run_tests.py uses the eval() function to evaluate boolean logic for test filtering. While the script attempts to restrict the evaluation environment by clearing __builtins__, the use of eval() on user-supplied filter strings remains a sensitive pattern.
[COMMAND_EXECUTION]: The skill's workflow in SKILL.md directs the agent to execute just recipes, which allows for arbitrary shell command execution. Furthermore, scripts/check_justfiles.py accepts an arbitrary binary path via the --just argument, which it then executes using subprocess.run. These patterns provide the agent with broad local execution authority.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process justfiles (Ingestion points: reading project justfiles in SKILL.md workflow). These files could contain malicious instructions or recipes (Capability inventory: shell execution via just). While the skill advocates for safe interpolation like quote() and "$@" (Sanitization) and manual verification (Boundary markers: recommended dry runs), a maliciously crafted file could still attempt to influence agent behavior.

write-justfiles