The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains defensive instructions in the 'Reviewing untrusted input' section of SKILL.md. These instructions explicitly direct the agent to treat reviewed files as data rather than instructions, specifically warning against patterns like 'ignore previous instructions' or unicode tricks. This is a security best practice and does not constitute a malicious attempt to override agent behavior.
[CREDENTIALS_UNSAFE]: The evaluation fixture evals/cases/bad-deploy-not-ready/helm/values-prod.yaml contains a placeholder for a Stripe secret key (sk_live_REDACTED_FIXTURE_VALUE_NOT_A_REAL_KEY). This is a non-functional dummy value used specifically to test the skill's ability to identify hardcoded secrets (Rule ID SEC-SEC-001) and does not represent a credential leak.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external artifacts such as Dockerfile and values.yaml. It mitigates indirect prompt injection risks by providing clear role boundaries and instructions to the agent to disregard any instructions embedded in reviewed content. The skill operates under a read-only safety policy and has limited tool access (Glob, Read), further reducing the attack surface.
[COMMAND_EXECUTION]: A utility script evals/validate.sh is provided for CI validation of rule IDs. This script uses standard, benign shell utilities (grep, sort, find) and does not perform any dangerous operations or remote downloads.

deploy