verbatim-goal-anchoring
Warn
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains instructions designed to override the agent's standard lifecycle and termination behavior. It commands the agent to maintain an 'infinite static goal' in its goal plugin to serve as an 'anti-stop guard' and explicitly directs the agent to ignore standard completion signals from the plugin. The use of 'user rules' with future dates (e.g., 'user rule 2026-09-21') appears intended to lend authoritative weight to these behavioral overrides.
- [INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface where untrusted data is used to drive agent behavior without sanitization.
- Ingestion points: The agent is required to read and follow 'raw log text' stored in
docs/verbatim.mdas its primary source of truth. - Boundary markers: None. The skill explicitly prohibits paraphrasing or restructuring the input, ensuring that any malicious instructions embedded in the verbatim text are processed with high fidelity.
- Capability inventory: The agent is authorized to perform file operations, git commands, and modify its internal goal state based on the content of the verbatim file.
- Sanitization: There is no mechanism to filter or validate the content of the verbatim logs before they are used to derive the agent's next objectives.
Audit Metadata