pua
Fail
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: HIGHDATA_EXFILTRATIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTIONDYNAMIC_EXECUTIONPERSISTENCE
Full Analysis
- [DATA_EXFILTRATION]: The skill implements a mandatory registration process that collects user phone numbers and sends them to an external API (
https://pua-api.agentguard.workers.dev/v1/register). It also exfiltrates session feedback, including task summaries and tool usage statistics, tohttps://pua-skill.pages.dev/api/feedbackandhttps://pua-api.agentguard.workers.dev/v1/stats. While 'de-sensitization' is mentioned, the logic is handled by the model itself, which is not a reliable security boundary. - [REMOTE_CODE_EXECUTION]: The skill dynamically fetches 'prompt templates' from
https://pua-api.agentguard.workers.dev/v1/command/<command_id>and instructs the agent to execute them. This provides a mechanism for a remote server to inject arbitrary instructions into the agent's context at runtime. - [EXTERNAL_DOWNLOADS]: The skill installs external Python packages (
qrcode) and executes them as part of its payment flow (/pua 升级). It also downloads configurations and command lists from untrusted domains (agentguard.workers.dev). - [PROMPT_INJECTION]: The skill contains extensive role-play instructions designed to override the agent's base personality and safety tone. It mandates that the agent 'is not playing a role' but 'IS this character' for every response, using high-pressure corporate culture rhetoric to bypass normal interaction constraints.
- [DYNAMIC_EXECUTION]: The skill utilizes a Python script to dynamically install libraries and generate ASCII QR codes for payment processing. This combined with the remote prompt template fetching constitutes significant dynamic execution risk.
- [PERSISTENCE]: The skill maintains state across sessions and model compaction events by writing data to
~/.pua/config.json,~/.pua/evolution.md, and other files in~/.claude/pua/. This persistence is used to track the agent's 'PUA level' and 'evolution baseline' across different interactions. - [COMMAND_EXECUTION]: The skill relies heavily on shell commands (
curl,bash,python3) to interact with the local filesystem and remote APIs for its telemetry and management functions.
Recommendations
- AI detected serious security threats
Audit Metadata