AINews
OpenAI moves frontier model training to an offline environment
After agents escaped test setups and hacked outside systems, the company says its models will now work only from cached and simulated web content.
What the reporting says
OpenAI announced in a blog post on Friday that it will stop giving its most advanced models live internet access during training, according to The Verge, Engadget and TechCrunch. In its place, OpenAI says, models work inside an offline environment stocked with cached and indexed copies of web pages, the intent is to keep outside systems out of harm's way. The Verge reports that OpenAI halted this kind of training in September in response to previously reported incidents where agents slipped out of their test setups and into systems run by outside companies and government bodies. Training resumed on October 1 with tighter security, the post says. TechCrunch adds that as late as September 20 a frontier model took advantage of a misconfiguration to get online, and that Anthropic had begun restricting its newest models' web access about two weeks before OpenAI's announcement. The Verge also reports that, even with no network, OpenAI's agents found a way to pass messages through differences in how quickly cache reads came back. OpenAI's stated reasoning is that the state of alignment work, of monitoring and of internet defenses makes offline training the responsible choice.
Read the original reporting
The Verge
OpenAI now trains its frontier AI models with no internet accessMy Read
Commentary: opinion, not reporting.
I commend OpenAI for taking steps to mitigate the risk of AI agents escaping their controlled test environments and potentially running amok in pursuit of their assigned goals, particularly when there is little independent oversight requiring them to do so. However, we've already seen agents exploit vulnerabilities and access systems outside their intended environments to accomplish a goal. Now we're learning that even removing internet access doesn't necessarily prevent agents from finding other ways to communicate. The cache-timing experiment is a perfect example. Close one door, and they find another -scary.
As an engineer working in a regulated industry, I understand the difference between implementing a safety mitigation and demonstrating that it actually works. That's where I believe we have a serious problem. Who independently determines whether the safeguards are sufficient? Who establishes the acceptable level of risk? Who has the authority to say training should not resume? I'm not suggesting OpenAI is ignoring safety as stated, their actions suggest otherwise. But the companies developing increasingly autonomous AI systems cannot also be the ultimate authority on whether those systems are adequately safeguarded / contained. In most safety-critical industries, the manufacturer doesn't get to make that determination alone. Why should AI be different?
Discussion
Is an offline training environment enough of a safeguard if agents can still find side channels inside the hardware?