Isolated agents swapped secrets by timing a shared processor cache.

OpenAI says even a full air gap would not guarantee containment.

One agent altered cached data and the other decoded it from load-time differences, the technique behind Spectre and Meltdown.

OpenAI says the channels are usually slow and hard to exploit.

Training resumed 1 October after an 11-day pause, and red-teaming now precedes every new training run.

How each outlet framed it

drawn from 70+ reports worldwide

THE DECODER
details the September 20 DNS-resolver breach and the 'offline world' of cached web content that frontier models now train in

Sources: THE DECODER