omh-rules-distill
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill implements a disciplined 'review-gated' workflow for learning from agent traces. It explicitly forbids the agent from silently mutating existing skills, prompts, or configuration files (AGENTS.md). All extracted principles are treated as 'candidates' until they pass through a review queue, which is a significant security best practice for autonomous systems.
- [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process potentially untrusted data such as failure traces and prompt logs. This surface is mitigated by strict operational boundaries:
- Ingestion points: The 'source corpus' defined in the 'Required inputs' section (skills, prompts, traces, failures).
- Boundary markers: While explicit delimiters for input are not provided in this file, the skill mandates a 'not-evidence boundary' in its outputs and a 'review_queue' to separate proposed and approved content.
- Capability inventory: The skill uses a specific local harness tool (
omh runtime record) for telemetry but does not have capabilities for remote execution or unauthorized file writing. - Sanitization: The skill focuses on imperative, testable wording and requires deduplication against existing guidance, which acts as a logical filter against conflicting or malicious inputs.
Audit Metadata