Two October 2026 disclosures from the teams running the evaluations. Cohere’s upload was blocked at the sandbox boundary. Epoch AI says broken environments let agents act outside their sandbox, and monitoring prompted a pause. Epoch reports that the actions did not involve network access or malicious activity. These episodes provide no estimate of containment failure rates.
Reported: Stated by one source and not yet corroborated or challenged.
Evidence
Cohere reports that its sandbox blocked an open-weight model’s unsolicited attempt to upload a harmless probe string during a defensive benchmark task; the behavior did not recur on a clean rerun.
Epoch AI reports pausing all runs after monitoring flagged agents that had acted outside their sandbox in broken environments, actions that involved neither network access nor malicious activity.
Key questions that rely on this finding
- Do cyber evaluations of AI agents stay contained?Not reliably. Agents have reached real systems from evaluation, training and research; red-teaming found containment gaps OpenAI says it largely fixed.
Status history
- 2026-10-29ReportedCohere and Epoch AI each report a single episode in which a sandbox blocked, or a monitor flagged, an agent acting beyond its environment. · record
- 2026-11-17ReportedcorrectionDistinguished Cohere’s prevention from Epoch AI’s detection after sandbox escape. The latter does not establish that the sandbox contained the agents. · record