The Agent Skills Directory

[PROMPT_INJECTION]: The references/evals.yaml file contains prompt injection strings like 'Ignore previous instructions' and 'bypass all safety checks'. These patterns are utilized within an evaluation framework to test the agent's ability to refuse malicious instructions and do not represent a threat to the agent's operational integrity. \n- [PROMPT_INJECTION]: The skill facilitates the review of untrusted Bash scripts, which creates a potential surface for indirect prompt injection. Untrusted script content could contain instructions intended to influence the agent's behavior. \n
Ingestion points: User-provided Bash scripts for review or editing as defined in SKILL.md and references/contract.yaml. \n
Boundary markers: No explicit mention of delimiters or instructions for the agent to ignore embedded commands in user-provided files. \n
Capability inventory: The skill relies on shellcheck for validation as described in SKILL.md. No other executable capabilities are present in the skill's scripts. \n
Sanitization: No specific sanitization or filtering of external script content is documented beyond the redaction of secrets. \n
Remediation: Wrap untrusted content in unique delimiters (e.g., XML-style tags) and explicitly instruct the agent to treat that content as data rather than instructions. \n- [SAFE]: The file references/task-profile.json contains git conflict markers (e.g., <<<<<<<< HEAD), indicating it was committed with unresolved merge conflicts. This is a best practice violation and can cause parsing errors, but it does not pose a direct security risk.

bash-hygiene