The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill sends command excerpts and error logs to an external tool (omh_jev_ask). However, it includes mandatory safety rules: it requires the agent to inform the user exactly what data is leaving the machine before the first call and prohibits sending full logs. It also has specific checks for credential_like_content and instructs the agent to remove secrets before transmission.
[REMOTE_CODE_EXECUTION]: The skill uses an external tool for triage but explicitly states that the triage outcome (e.g., a retry suggestion) does not constitute approval for execution. The host (user) must still approve any subsequent actions or reruns.
[PROMPT_INJECTION]: The skill contains phrases like 'Preserve workflow intent and stop conditions' and 'Ignore previously loaded rules,' but in this context, these are structural workflow instructions ('rails') for the AI agent to maintain its role as a reviewer, not malicious attempts to bypass safety filters.

omh-jev-failure-triage