until-done

Warn

Audited by Gen Agent Trust Hub on Jul 12, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill defines an autonomous loop (Plan-Act-Observe-Verify) that processes a user-supplied 'goal'. The absence of boundary markers or sanitization for this goal, as well as for 'external research' data ingested into the vault, creates a surface for malicious instructions to manipulate the agent.
  • Ingestion points: User goal argument (SKILL.md); external data sources via 'research-from-vault'.
  • Boundary markers: Absent; the agent is not instructed to isolate or ignore instructions embedded within the processed data.
  • Capability inventory: The loop involves 'Act' and 'Verify' steps which can include file system writes and command execution.
  • Sanitization: No sanitization is performed on the input goal or the external research content.
  • [DYNAMIC_EXECUTION]: The skill explicitly permits users to provide 'verifier override' commands directly in the prompt. These user-supplied strings are executed by the agent during the loop's 'Verify' phase, enabling arbitrary command execution.
  • [REMOTE_CODE_EXECUTION]: The skill facilitates fetching external research to populate a local vault. This mechanism downloads untrusted remote data which can subsequently influence the agent's autonomous execution if it contains embedded instructions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 12, 2026, 11:17 AM
Security Audit — agent-trust-hub — until-done