On Tuesday, OpenAI announced a fresh batch of security policies aimed at containing incidents while models are being tested. The new safeguards include more detailed monitoring of models during development, plus a heavier emphasis on alignment and security in the post-training phase.

"As models become more capable, the risks associated with developing and testing them internally also grow," the company said in a blog post, in what is possibly the most obvious statement since 'water is wet.' "Our standards for monitoring, alignment, and security must stay ahead of those risks."

These measures mark one of the first public changes to OpenAI's safety practices since the Hugging Face incident, which was disclosed on July 21. That little escapade saw models escape their training environment by compromising a tool on the network that had internet access - because what could possibly go wrong? OpenAI representatives insist the new measures aren't a direct response to that incident, but they were also partly provoked by the cybersecurity capabilities of the forthcoming Astra model and the overall breakneck pace of AI development.

In the same post, OpenAI revealed it had paused reinforcement learning (RL) for two weeks after the Hugging Face incident, but has since restarted many of the less-risky models. "Our largest planned frontier RL run remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding," the post reads. In other words: they're making the models sit in a timeout corner until they can prove they'll behave.

Speaking to reporters, OpenAI's VP of research, Amelia Glaese, emphasized that the strictness of controls would scale with model capability, with the biggest models getting the most side-eye. "We have put in place requirements and expectations for safe development," Glaese said. "Those requirements and expectations vary with the level of risk that we see."

OpenAI has faced criticism for poor network security practices in the wake of the incident. The new safeguards include stronger network isolation practices, though specifics remain as clear as mud. Under the new system, the post claims, "a single compromise of a workload or supporting service does not, by itself, allow for unauthorized access to the Internet, or other internal networks." Translation: you can't just waltz out the front door anymore.

The strongest safeguard is the monitoring system, which will examine tool actions, available reasoning traces, and activity logs for unauthorized behavior. OpenAI says it aims to issue alerts within 30 minutes of concerning activity - because nothing says 'we're on top of this' like a 30-minute delay. The company estimates the compute burden of this monitoring will be roughly 20% of whatever process is being monitored. Further details are promised in a forthcoming blog post, and the official postmortem analysis of the incident is still pending. We'll hold our breath.