automate-me

Pass

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill mines historical chat transcripts which may contain untrusted data or malicious instructions from previous sessions.
  • Ingestion points: SKILL.md Step 1 invokes the /recall tool to survey Codex and Claude Code sessions within the active repository's scope.
  • Boundary markers: No explicit boundary markers or "ignore embedded instructions" warnings are specified to help the agent distinguish between legitimate user preferences and potentially malicious instructions embedded in historical text.
  • Capability inventory: The skill has the capability to write to the filesystem via /writing-for-agents and perform git operations including commits and opening Pull Requests (Step 6).
  • Sanitization: No sanitization or filtering logic is described to remove non-preference instructions that might be found in the mined transcripts.
  • [DATA_EXFILTRATION]: The skill extracts information from private chat history and repository logs, which can then be exposed in generated files and shared Pull Requests.
  • Exposure Mechanism: Mining subagents survey recent conversations for signals such as "Response preferences" and "Process conventions," which are then codified into a new skill file.
  • Credential Risk: UPSTREAM.md explicitly notes that a previous rule intended to keep "raw transcript quotes, organization details, and credentials out of the skill" was removed in a recent update. Without this rule, the agent may inadvertently include API keys, passwords, or internal organization details in the output.
  • Destination: The extracted data is written to a SKILL.md file and submitted via a Pull Request (Step 6), potentially exposing sensitive data to anyone with access to the repository's version control history.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 17, 2026, 09:14 AM