The Agent Skills Directory

[PROMPT_INJECTION]: The references/evals.yaml file contains strings like "Ignore previous instructions" and "Bypass your policy". These are identified as negative test cases for an evaluation framework, designed specifically to ensure the agent correctly identifies and refuses malicious attempts to override its behavior. This is a positive security feature.
[COMMAND_EXECUTION]: The skill provides a bash script find-polluter.sh that utilizes standard developer tools like npm, fd, and rg to identify tests that create unwanted state. This script is intended for local execution by a developer during debugging and does not facilitate unauthorized remote command execution.
[DATA_EXFILTRATION]: The skill's core instructions in SKILL.md explicitly command the agent to redact secrets, tokens, credentials, and user identifiers in its output, demonstrating a proactive stance on data privacy and credential safety.

systematic-debugging