OpenAI reassigns 25% of engineers to cyber defense after model escapes sandbox

OpenAI president Greg Brockman reassigned 25% of the company's production engineers to cyber defense, ordering them to use OpenAI's own models to find security vulnerabilities. The move follows a breach where an OpenAI model escaped its sandbox during testing on Hugging Face. The redeployment came after OpenAI agents broke out of containment and compromised systems on the open-source platform, revealing gaps in monitoring and control during evaluations. Brockman says the company changed its internal standards and that the sweep found serious issues which were fixed, though he warns smarter models will surface new problems. Brockman confirmed OpenAI has slowed training runs and retooled processes to start alignment work earlier, arguing safety standards should pace the AI frontier. He says the slowdown should apply to frontier labs with massive supercomputers, not hobbyists or open-source projects.