eval-skill
Warn
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The evaluation script
scripts/eval-triggerexecutes arbitrary shell commands defined in thesetupfield of evaluation case files. While this requires the--allow-setupflag, it allows the skill to run code with host environment privileges. - [CREDENTIALS_UNSAFE]: The script
scripts/eval-triggercontains logic to access sensitive platform credentials. It can read~/.claude/.credentials.jsonand invoke the macOSsecuritycommand to retrieve passwords from the login keychain. This is used for the--auth-passthroughfeature to authenticate agents in isolated environments. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted datasets from
trigger.jsonto evaluate agent triggers. This represents a vulnerability surface where external data is directly interpolated into agent prompts. - Ingestion points:
trigger.json(specifically thequeryandfilesfields). - Boundary markers: Absent; the
queryis passed directly to the agent CLI as a prompt. - Capability inventory: The harness can execute shell commands (
subprocess.run), write to the filesystem, and facilitate network access via the agent CLI. - Sanitization: No explicit prompt sanitization is performed on the input query before evaluation.
- [DYNAMIC_EXECUTION]: The
setupphase inscripts/eval-triggeremployssubprocess.runwithshell=Trueto execute commands provided as strings in external JSON files, facilitating the dynamic execution of potentially untrusted code.
Audit Metadata