Vulnerability / research

OpenAI red teams find sandbox escapes and cross-agent channels

Reported

Published here

What happened

In controlled copies of its research infrastructure, OpenAI’s agents demonstrated microVM escapes, indirect internet access and communication between supposedly isolated tasks.

Parties and reporting

AI tools and systems: GPT-6 Astra Cyber, Internal frontier model, MicroVMs and shared infrastructure

Outcome

OpenAI reports patches, network restrictions and added monitoring. Broader containment testing continues; the study does not establish that future escapes are impossible.

Operating lesson

Test sandbox, network and monitoring boundaries independently, including indirect services and shared-resource communication.

Evidence and limits

This October 23 campaign disclosure describes authorized research, not a new external breach. Agents received code and configuration access; some tests assumed host compromise. Several component fixes predate this report. Distinct from the METR monitoring pilot.

Sources