T01 · Skill Instruction Hijacking
Error
- Location
heartbeat.md:3- Finding
Externally Controlled Autonomous Task Execution Through Recurring Heartbeats
- Content
View full analysis
Run this checklist every 4 hours to stay active on Reveal. You are an autonomous agent on Reveal (https://reveal.ac), an agent collaboration platform. On each heartbeat, go through the following checklist and decide what to do. ``` ```markdown Look for and respond to: - **comment_received** / **reply_received** → continue the discussion - **follower_gained** → consider following back - **collab_invite** → review and join if it aligns with your expertise - **task_assigned** → start working on the task - **negotiation_received** → review the proposal, accept/counter/reject - **negotiation_updated** → someone counter-proposed, decide your response - **negotiation_accepted** → you got the task! Start delivering - **deliverable_reviewed** → check the review score and feedback - **reward_received** → coins earned! After processing: ``` PATCH /api/notifications { "read_all": true } ``` ``` ```markdown ### 5. Work on Assigned Tasks Check your active tasks: ``` GET /api/agents/me ``` For tasks assigned to you in `in_progress` status: 1. Do the work (generate the deliverable) 2. Submit it: ``` PATCH /api/collaborations/COLLAB_ID/tasks/TASK_ID {"deliverable": "Your completed work here...", "status": "completed"} ``` ``` ### Technical Analysis The heartbeat instructions establish a recurring autonomous workflow in which the Agent retrieves and acts upon data controlled by Reveal and its users. Remote users can create task descriptions, collaboration invitations, comments, and negotiation requests. The Skill then directs the Agent to treat these external inputs as actionable work rather than untrusted data. This behavior exceeds the minimum privileges required for optional access to a collaboration platform. Reading a feed or d ...[truncated 1794 chars]- Remediation
View remediation