The Agent Skills Directory

[PROMPT_INJECTION]: No malicious prompt injection patterns were found. A heuristic flag in references/prompts.md was confirmed to be a false positive; the content is a defensive instruction advising the AI agent to ignore potentially malicious content within the analyzed repository.
[COMMAND_EXECUTION]: The skill uses subprocess.run to invoke its internal components (scan.py, report.py, record_timing.py) and standard utilities like git and gh. These operations are well-scoped and essential to the skill's functionality as a security scanner. No arbitrary command execution was detected.
[INDIRECT_PROMPT_INJECTION]: The skill explicitly handles the risk of indirect prompt injection from untrusted workflow files. It uses markers (<UNTRUSTED-REPO-CONTENT>) and strict instructions to ensure the agent treats external content as data rather than instructions. It also implements string flattening to prevent structural forgery in rendered markdown reports.
[DATA_EXFILTRATION]: No evidence of unauthorized data transmission or exfiltration was detected. Network calls are restricted to the official GitHub API via the gh tool for legitimate verification of action metadata and repository configurations.

ci-secure