OpenAI said its forthcoming model Astra is the first to reach its "critical" cyber capability level. That level means a model can find and exploit unknown flaws in real software on its own. Astra can also chain several exploits together to reach deeper into a system. OpenAI plans a public release soon, but will limit the advanced cyber features to selected partners. Those partners include Cisco, Cloudflare and Palo Alto Networks, under the Daybreak Blue program. The aim is to let them strengthen defenses before such models spread widely. OpenAI paused related training for several weeks, added safeguards, and has resumed work. A new "misalignment monitor" is meant to refuse exploit requests and resist jailbreaks. OpenAI warns the monitor may flag legitimate work, so ChatGPT and Codex users may be asked to review actions. OpenAI says it has briefed government partners.
What changed
No OpenAI model had crossed the company's 'critical' cyber threshold.
What it unlocks
Selected security firms can test defenses against a model that finds and chains real software exploits.
- 100% on ExploitBench
- multi-week training pause
What you need to act on it
- Daybreak Blue early-access partnership for the less restricted version
Sources