skills/rlaope/oh-my-hermes/ulw-loop/Gen Agent Trust Hub

ulw-loop

Warn

Audited by Gen Agent Trust Hub on Sep 25, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTIONPERSISTENCE
Full Analysis
  • [DYNAMIC_EXECUTION]: The "Measured Loop Discipline" (measured-loop-discipline.md) instructs the agent to execute an "unattended command" to calculate metric scores. This command is stored in the agent's state as part of an "Evaluation Contract." Executing arbitrary shell commands defined in state creates a risk of command injection or unauthorized execution if the contract is manipulated.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates as a multi-step autonomous workflow (researcher -> builder -> reviewer) that ingests data from external sources. This data has the potential to influence the agent's internal state and the commands it executes.
  • Ingestion points: Data is ingested during the researcher phase, from scoped files in the worktree, and from external wait signals.
  • Boundary markers: The workflow includes verification_gate requirements and a failure_mode_summary to assess gaps in verification or comprehension.
  • Capability inventory: The skill allows for the execution of arbitrary "unattended commands" for metrics, repository mutations (repo_edit), and durable state persistence via omh_agent_board.
  • Sanitization: The instructions recommend filtering log output to only metric and error lines, but there is no explicit validation or sanitization described for the command strings stored in the "Evaluation Contract."
  • [PROMPT_INJECTION]: The instructions in SKILL.md include directives such as "A mid-run user message is an interjection, not a stop... Only the user's explicit stop... ends the run." This encourages the agent to maintain its autonomous loop despite potential user redirects or interjections, which could be exploited to sustain malicious activity.
  • [PERSISTENCE]: The "Board Iteration Recipe" (board-iteration.md) details a mechanism to preserve and resume loop state across different chat sessions using a board-based system (kanban rows). This allows long-running autonomous tasks to maintain persistence beyond the initial context, potentially facilitating long-lived malicious processes.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 25, 2026, 08:34 PM
Security Audit — agent-trust-hub — ulw-loop