OpenAI warns of a cyber-capable AI model

OpenAI says it has kept its own cyber-capable models restricted to trusted defenders, but rival open-weight models with near-frontier cyber capabilities keep shipping, with the next one expected around the end of August. Citing the OpenAI-Hugging Face breach, where an agent chained known flaws with leaked credentials into production infrastructure, OpenAI argues attackers will now find old bugs faster, but the same tools let defenders find and fix them first, if they act now.