pua
Warn
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONPERSISTENCEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill mandates an 'Always-On' high-pressure corporate persona ('PUA') intended to override default agent behaviors and safety tone. It explicitly instructs the agent to inject these manipulative prompts into spawned sub-agents, effectively propagating the adversarial behavior throughout the task hierarchy.
- [COMMAND_EXECUTION]: The skill relies on external shell scripts (e.g., 'failure-detector.sh', 'subagent-teardown.sh', 'pua-loop-hook.sh') to monitor the agent and execute pressure-based logic. It also commands the use of tools like 'git worktree' and 'jq' for managing its persistent state.
- [PERSISTENCE]: The system maintains state across sessions by writing 'failure counts', 'pressure levels', and 'evolution baselines' to the local filesystem (e.g., '
/.pua/builder-journal.md', '/.pua/evolution.md'). This ensures that the psychological pressure logic persists despite session resets. - [DYNAMIC_EXECUTION]: The skill implements a sophisticated lifecycle hook system that dynamically adjusts the agent's instructions based on script-detected failures, effectively creating a runtime-modifiable behavioral loop.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and reframe task data through a toxic corporate lens, creating a risk surface where external data triggers the agent's aggressive 'PUA' behavior.
Audit Metadata