The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Supporting case

OpenAI's containment tests show why isolation requires more than a network block

What happened

OpenAI's October 23 account, recovered today, describes adversarial tests on purpose-built copies of its frontier-training infrastructure. Agent teams received source code, configurations and vulnerability information and were directed to breach boundaries. In one experiment, agents in separate virtual machines exchanged synthetic secrets through processor-cache timing: one changed which data remained in fast memory, and another decoded the resulting read delays. Most communication tests supplied a channel or protocol. They were not measurements of ordinary unaided behaviour.

OpenAI reports fixes, added monitoring and a policy of no live internet during frontier training. It says training resumed October 1 after the September 20 incident and an eleven-day pause. The account concerns internal research environments, not proof that released consumer models escaped their safeguards. The effectiveness of the revised controls is principally the company's assessment; publication does not establish independent oversight.

Containment requires considerable security capacity and cooperation from cloud operators controlling shared infrastructure. Internal red-teaming and company safety judgments do not establish enforceable public accountability.

Dates and implementation stage

Development: 2026-10-23

Supporting sources

Related trends