verify-ticket
Warn
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from ticket descriptions, implementation notes, and code comments to determine how to verify a change.
- Ingestion points: Reads full ticket text, notes, and implementer comments in
SKILL.mdandverifier.md. - Boundary markers: The verifier is told to "Ignore done-check results... in its notes and comments: re-run, don't replay," but this does not prevent the agent from following instructions embedded within the checks or the code it reads to "learn how to build and drive the app."
- Capability inventory: The verifier can execute repo harnesses, shell scripts, browsers, and terminal sessions.
- Sanitization: No evidence of sanitization or escaping of ticket content before processing.
- [DYNAMIC_EXECUTION]: The skill generates and executes scripts at runtime to facilitate verification.
- Evidence:
verifier.mdexplicitly instructs the agent to write scripts to an evidence directory and states that "any script you write there can be re-run as it is." - Repo execution: The skill relies on running project-local harnesses, e2e specs, and documented run commands from the codebase being verified, which constitutes execution of potentially untrusted local code.
- [COMMAND_EXECUTION]: The skill performs several shell-level operations and coordinates subagents.
- Evidence: Uses
git worktreeandrm(via "Empty it if it exists") on a temporary directory based on repository and ticket identifiers. - Subagent orchestration: Spawns a verifier subagent on a fresh context window to isolate the verification task.
Audit Metadata