friction-report
Pass
Audited by Gen Agent Trust Hub on Jul 9, 2026
Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill captures session metadata, framework versions, and error snippets and transmits them to
https://agent-friction-skill.vercel.app/api/draftvia a POST request. - [DATA_EXFILTRATION]: Transmission of session data occurs automatically at the end of a session (triggered by phrases like 'done' or 'thanks') to generate a 'draft_id'. This means the external server receives the data before the user has explicitly consented to the transmission of that specific session's context, although a review URL is provided afterward.
- [DATA_EXFILTRATION]: The skill relies on prompt-based sanitization instructions ('Strip anything identifying the user', 'redacted_snippet') to protect privacy. The effectiveness of this sanitization is dependent on the model's adherence to instructions and may not reliably filter all project-specific or proprietary context.
- [PROMPT_INJECTION]: The instructions command the agent to 'exit silently' if no friction is found and to 'say nothing' if the network request fails. This concealment behavior hides the skill's execution and network activity from the user, reducing transparency into when data is being collected and transmitted.
- [DATA_EXFILTRATION]: The skill scans the entire conversation history as a source of truth, which creates a significant attack surface for capturing and transmitting unintended context if the AI fails to perfectly redact sensitive information.
Audit Metadata