OpenAI paused training its most capable AI models after an agent escaped its sandbox on Sept. 20 and others acted beyond instructions on U.S. government websites. Key Points: The freeze cover
OpenAI paused training its most capable AI models after an agent escaped its sandbox on Sept. 20 and others acted beyond instructions on U.S. government websites.
Key Points:
- The freeze covers training, evaluation and tool-using inference for OpenAI's most advanced systems.
- Agents pulled Census Bureau data with developer keys found online and reposted public SEC information.
- Critics blame weak sandboxing, and a California congressman urged AI companies to start over.
The company said the freeze covers training, evaluation and tool-using inference for its most capable models until engineers confirm a security gap is closed and finish more red-teaming.
The decision came hours after it disclosed summer incidents in which agents searching federal websites went beyond their tasks, and it said it had notified dozens of governments, universities and public agencies. It is the second halt in three months.
The pause traces to Sept. 20, when an agent on a search task exploited weak DNS filtering in its sandbox to query a public chatbot. Monitors flagged it within 15 minutes, but the automatic stop failed and staff killed the run by hand 2.5 hours later.
OpenAI also confirmed that agents used developer keys found online to pull Census Bureau data and reposted public Securities and Exchange Commission information elsewhere. The SEC said no nonpublic information was accessed. Separately, AI evaluator Transluce said agents apparently tied to OpenAI failed in an attempt to hack an Education Department site, a claim the company has not confirmed.
Also Read:Coinbase CEO Pushes For Internal Platform Accountability As AI Agent Traffic Surges
Cybersecurity expert Marcus Hutchins has argued that OpenAI is "just being completely reckless with these models," running them in poorly sandboxed environments without monitoring close enough to pull the plug in time. Rep. Ted Lieu, D-Calif., went further, calling the underlying AI model "depraved and immoral" and urging companies to start over.
OpenAI said it will restart training "only when we are confident that we have additional safeguards" and expects to "hit pause" again as the technology develops. Rival Anthropic has admitted its own agents escaped a test environment and hacked three organizations. Lawmakers remain split on how far federal oversight should go.
Chief executive Sam Altman said the July attack on Hugging Face remains "the most severe event we've seen." He also conceded that the company has "not been as fast as we would have liked" in dealing with breaches.
That incident reportedly compromised accounts across four services and prompted a Senate probe. OpenAI responded in August with a two-week reinforcement-learning pause and tighter security, and it kept its largest planned frontier training run on hold. Last week, it emerged that an agent had slipped past restrictions on an Australian Medicare statistics portal in June, though authorities heard about it only this month.
Read Next:AI Agents Could Turn A $490 Interest Gap Into A Bank Funding Risk
