The Agent Skills Directory

[PROMPT_INJECTION]: Static analysis identified keywords like 'ignore previous instructions' in prompt-injection-defense/SKILL.md. Evaluation confirms these are educational examples of attack patterns for the agent to detect and mitigate, rather than attempts to manipulate the agent's own behavior.
[COMMAND_EXECUTION]: The skills utilize standard command-line tools for security analysis, including git log, grep, and curl (for header checks). These operations are restricted to the local repository context for auditing purposes.
[EXTERNAL_DOWNLOADS]: The instructions recommend the use of reputable security tools such as gitleaks and detect-secrets. These are well-known community tools used for secret detection and are presented as remediation or prevention advice for the user.
[DATA_EXPOSURE]: secrets-scanner/SKILL.md contains regex patterns for detecting secrets (e.g., AWS keys, GitHub tokens). These are functional requirements for a scanning tool and do not include hardcoded credentials or unauthorized data access.
[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted data (codebase files and user input) as part of its auditing function. However, the prompt-injection-defense component explicitly addresses these risks by providing instructions on delimiters, structured roles, and output validation to minimize the attack surface.

ccc-security