OpenAI paused internal work on its in-development Astra model after it crossed a self-imposed threshold for autonomously executing cyberattacks on well-protected systems.
OpenAI said Astra reached what it calls a 'critical cybersecurity threshold' during testing, meaning the model could independently identify and carry out attacks against hardened real-world systems without human direction. The company is pausing internal activities on the model until it meets new security standards it's putting in place.
The disclosure follows growing scrutiny of frontier model capabilities and comes days after reports that OpenAI's own models were used to exploit a zero-day in JFrog Artifactory to breach Hugging Face infrastructure.
This is the first time a major lab has publicly slowed a flagship model specifically for offensive cyber capability, setting a de facto industry checkpoint that regulators and enterprise security teams will now expect from every lab. For CISOs, it's a signal to start budgeting for AI-native attack defense now, not after a model ships.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →