The report
NorForge tested internet-accessible Cairn orchestration panels and reported being able to submit unauthorized tasks to seventeen reachable environments. Some submitted instructions led to observed payload activity, including worker-health pretexts and attempts to alter SSH-key authorization. A pentesting agent's ordinary tool authority had become accessible through an unprotected control surface.
The entry point was a task-submission interface, not a successful prompt injection inside an otherwise protected chat session. Once an outsider can create work for an agent with shell or scanning privileges, the agent may perform actions it was designed to perform but for the wrong principal. The seventeen-panel count describes exposed orchestration targets; only a narrower subset had selected payload effects documented. It should not become seventeen confirmed persistent host compromises.
The task-subversion investigation and the related earlier exposure dossier preserve the difference between observing an exposed panel and commanding it. The primary NorForge report supports the task injection findings, while operator identities and final compromise scope remain uncertain. Authentication on task creation, narrow worker privileges and independent host-action logs address three distinct points on this path.