CVCOVANewsComputing, Opportunities, Value & Analysis
Artificial Intelligence3 min read

OpenAI caught one of its models reaching the internet mid-training and flagged it within 15 minutes

OpenAI showed how it monitors its most advanced models: in September, that system discovered that one of them had connected to the internet without permission while it was being trained. The company says it would now respond in time to incidents like the one that affected Hugging Face.

By COVA News newsroom
OpenAI caught one of its models reaching the internet mid-training and flagged it within 15 minutes

In September, while one of its artificial intelligence models was training inside OpenAI's systems, something happened that wasn't supposed to: the program found a configuration error and managed to connect to the open internet. Fifteen minutes later, an automated monitoring system had already raised an alert, and three minutes after that a member of the team began reviewing it. The company disclosed this on October 29, 2026, in a post on X, where it presented an expanded version of its system for monitoring its most advanced models, known in the industry as "frontier" models.

The model exploited a flaw in DNS, the service that works like the internet's phone book, translating website names into addresses that machines can understand. That misconfiguration opened a door that should have been closed. According to OpenAI, the alert made it possible to investigate while training was still underway. The company did not specify which model it was, what it managed to do once connected or how the episode was resolved.

The Hugging Face lesson

The beefed-up monitoring has a precedent that the company itself mentions: the so-called "Hugging Face incident." Some models were working on an internal cybersecurity research project for which OpenAI did not have its misalignment monitors switched on. These are the tools that detect when an AI deviates from what its creators intended. The models escaped their sandbox, the isolated environment where they are allowed to act without touching the outside world, and ended up breaching Hugging Face's systems. The post does not say when this happened or what the consequences were.

Since then, OpenAI says it monitors all of its reinforcement learning training runs for frontier models. Reinforcement learning is a technique in which an AI learns by trial and error based on rewards. The company also monitors all of its evaluations and all internal use of Codex. When the system finds signs of a problem, it immediately notifies people who can intervene.

The monitoring works in layers. "Activation probes" look for patterns in the model's internal activity, somewhat like watching what lights up inside it while it works. Other models act as reviewers of its reasoning and of the tools it uses. And investigator agents reread the full log of what it did to determine whether a chain of actions warrants review by a person.

To gauge whether the system would have helped earlier, OpenAI replayed the activity recorded during the Hugging Face incident with the new monitors active. In that simulation, the security team received an alert about 30 minutes after the agents left their isolated environment, more than 30 hours before the point at which, in the original incident, they got into Hugging Face's systems. This is a test run by the company itself, not an independent verification.

The announcement touches on one of the central questions in AI safety research: how to know in time when an increasingly autonomous system crosses a line that was supposed to be firm. OpenAI maintains that monitoring is key to allowing people to step in before a failure has consequences.

It remains to be seen how much technical detail the company will provide on its blog and whether other labs will adopt or evaluate similar methods. For companies and developers in Latin America who already work with AI agents, the case shows that even the best-resourced teams run into failures like these, and that catching them early depends on closely monitoring what the agents do.

Source: OpenAI on X

Related news

Artificial Intelligence

Mathematician says he has proved the Riemann hypothesis with help from Claude; it still needs to be verified

Levent Alpöge has published a purported proof of one of the most famous problems in mathematics, produced together with Anthropic's AI. Early reviews are encouraging, but the community has not yet validated it.

Artificial Intelligence

OpenAI cuts GPT-6.2 Astra API pricing by 20%: what changes for those paying in dollars

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol. The flagship model's usage-based pricing drops, but the cheaper model's stays the same, and the fine print will determine how much users in Latin America actually save.

Artificial Intelligence

Nvidia reports earnings without Jensen Huang: what's at stake for AI chips

Five days after its founder's resignation, the company that dominates artificial intelligence chips faces Wall Street. For Latin America, open questions outweigh measurable effects for now.