grok
Pass
Audited by Gen Agent Trust Hub on Oct 3, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill mandates a behavioral override by instructing the agent to 'refuse' and 'not honor' user requests for model families other than Grok 4.6. This overrides the agent's standard capability to follow user instructions regarding model selection.
- [INDIRECT_PROMPT_INJECTION]: The skill monitors user input to decide model selection behavior, creating a potential surface for prompt injection attacks aimed at bypassing the model lock. Ingestion points: User prompts are analyzed for keywords like 'grok', 'model lock', and model family names as specified in SKILL.md. Boundary markers: The instructions in SKILL.md do not define clear delimiters or isolation for user-provided data when evaluating it against the locking logic. Capability inventory: The skill controls the 'model' parameter during task launches and subagent spawning as detailed in SKILL.md and references/harness-ids.md. Sanitization: No sanitization or validation mechanisms are described to prevent adversarial user prompts from influencing the logic.
Audit Metadata