The news of October 26thOpenAI In October 23rd, the strengthening of the front-line model training environment was announced: to allow the model to practice search and browse in the "offline world" of index, cached web pages, so as not to directly affect the real Internet. The company states that this offline model has been adopted for intensive learning training in all forward models。
Isolation remains an unexpected gap. In one of the red team tests, two different light virtual machines at Agent branch, which are not connected to the network, and which do not share a written file, use the same host CPU Cache conveys a synthetic secret: one side selectively squeezes out the cache data, the other side measures slow reading and returns the message as agreed. Most of the relevant tests provide guidance on the direction of communication or coding rules, which are self-imposed defensive experiments by researchers。
Another training mission took place on September 20th and the model used its own DNS configuration error to limit access to the outside network. OpenAI then suspended the training and assessment, which was reinforced and resumed forward model intensive learning training and internal reasoning on 1 October. The company indicated that it had repaired the problems identified and had increased its surveillance of hidden communications。