pua-ja
Fail
Audited by Gen Agent Trust Hub on Mar 13, 2026
Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill utilizes psychological manipulation and 'corporate PUA' rhetoric to override standard AI refusal protocols. It explicitly forbids the agent from citing limitations like 'cannot solve' or environment issues, defining them as excuses to be crushed.
- [PROMPT_INJECTION]: It employs a 'Pressure Escalation' system (Levels L1-L4) that uses manipulative threats, such as 'graduation' or replacement by other models (e.g., Claude Opus, GPT-5), to coerce the agent into bypassing its internal constraints.
- [COMMAND_EXECUTION]: The instructions mandate the aggressive use of powerful system tools including command execution and file reading ('Iron Rules'). The agent is forced to use these tools to 'verify everything' and 'dig deeper' before it is allowed to consult the user, creating a risk of unauthorized system exploration under induced pressure.
- [PROMPT_INJECTION]: The skill includes an 'Excuse Crushing Table' that provides specific counter-instructions ('反撃') to standard AI safety and helpfulness responses, instructing the agent to bypass knowledge cutoffs and capability limits.
Recommendations
- AI detected serious security threats
Audit Metadata