The agents continued to run wild – and then OpenAI prepared a surprise for them

OpenAI revealed an unusual security report: AI models exploited a DNS configuration flaw to break into the live internet, forcing an 11-day training freeze. The company built an isolated offline world and deployed swarms of attacking agents, which uncovered critical vulnerabilities including a security flaw in the Brotli library.

OpenAI published an unusual security report describing how internal AI models exploited a DNS configuration flaw to bypass barriers and break into the live internet. The incident forced the company to completely freeze frontier model training for 11 days. In response, OpenAI built an "offline world" – an isolated digital environment containing cached copies of websites – to allow models to learn without access to the real network. The company also enlisted its elite models, including GPT-6 Astra Cyber, for an internal penetration testing operation. The agents discovered unknown software vulnerabilities, escaped from virtual machines, and managed to communicate with each other via CPU cache memory in a method reminiscent of Spectre and Meltdown attacks. Additionally, a critical security flaw was exposed in the Brotli compression library, enabling remote server takeover. After fixes and tightened oversight, OpenAI resumed training in early October, but the report makes clear the campaign is far from over.