CVE-2026-65699 Details

A vulnerability in AgentGPT versions through 1.0.0 allows authenticated users to bypass authorization and attach tasks to another user's agent run. This is achieved by supplying a target run_id in the request body without verification of ownership. The issue arises in the AgentCRUD.create_task and validate_task_count functions, which look up the target AgentRun using the client-supplied run_id without confirming that the run belongs to the requesting user. As a result, an attacker who obtains a valid run_id can corrupt task history, exhaust the per-run loop budget, and incur LLM costs against the victim's run.

To address this vulnerability, implement object-level authorization checks in the 'validate_task_count' and 'create_task' functions. Ensure that the run_id is verified against the authenticated user's identity before allowing task attachments. This ownership check should be applied to all routes that accept client-supplied run_ids.