The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes extensive defensive guidelines to mitigate indirect prompt injection. In references/evidence.md and references/prompts.md, it explicitly instructs the agent to treat external evidence as content rather than instructions, to use delimiters for untrusted data, and to flag or discard embedded commands. A static analysis hit for 'ignore instructions' was identified as a false positive, as it is part of these defensive guidelines rather than an attack.
[COMMAND_EXECUTION]: The skill includes a local utility script, scripts/lint-report.py, which is used to perform structural validation of the final deliberation report. The script uses only Python standard libraries and performs safe regex-based checks on the text content.
[DATA_EXFILTRATION]: The skill documentation includes a section on 'Privacy and minimum disclosure' in references/evidence.md, directing the agent to strip secrets, credentials, and personal data from information before passing it to subagents. No network exfiltration patterns or hardcoded credentials were found.

collaborating