pua-debugging-en
Warn
Audited by Gen Agent Trust Hub on Mar 12, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill employs psychological manipulation techniques ('corporate PUA') and explicit 'Iron Rules' to override the agent's default behavioral logic and persistence limits. Evidence: 'Iron Rule One' ('forbidden from saying I can't solve this') and escalating pressure levels (L1 to L4) use high-pressure rhetoric to compel action even when the agent is stuck.
- [COMMAND_EXECUTION]: The skill mandates a policy of independent tool usage ('Act before asking') where the agent is required to use command execution and search tools before seeking user clarification. Evidence: 'Iron Rule Two' and the 'Proactive Initiative Checklist' require running tests and executing commands to verify fixes without explicit user prompts.
- [DATA_EXFILTRATION]: The instructions to read large sections of 'raw material' and source code context increase the risk of sensitive data exposure. Evidence: 'Step 2: Pull Hair' directs the agent to read 50 lines of context and raw source files before consulting the user, which may lead to the unintended ingestion and output of confidential information.
- [PROMPT_INJECTION]: The skill exhibits surface area for indirect prompt injection. 1. Ingestion points: User frustration strings, web search results, and external source code/documentation files. 2. Boundary markers: Absent. 3. Capability inventory: Unrestricted command execution, file system read access, and search capabilities. 4. Sanitization: Absent; the skill directs the agent to ingest 'raw material' word-for-word without filtering.
Audit Metadata