hillclimb-skill
Pass
Audited by Gen Agent Trust Hub on Nov 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because its core functionality involves reading and acting upon data generated by external evaluations of other skills.
- Ingestion points: The
climb.py failuresutility readsrun.jsonfiles which contain the full text of task prompts, expectations, and execution transcripts from the evaluation harness. - Boundary markers: The skill instructions provide a manual boundary by telling the agent: "Wording from a task's prompt, expect:, assert or transcript stays out of the skill."
- Capability inventory: The skill possesses
Bash,Write, andEditcapabilities, which are necessary for modifying instructions and running evaluation scripts. - Sanitization: The skill lacks automated sanitization or filtering of the transcript content before it is presented to the agent for analysis.
Audit Metadata