OpenAI Unveils More Capable Cybersecurity Model With Tighter Safeguards

OpenAI is beginning to offer GPT-6 Cyber to a limited number of organizations doing authorized security work. The company plans to extend access to all customers in its Daybreak Red program over the coming weeks. It is also widening early access to GPT-6 Astra for defensive work through Daybreak Blue.

OpenAI says the new model exceeded its general-purpose GPT-6 Astra on most cybersecurity tests. It scored 48.5 percent on ExploitGym, compared with 42.4 percent for Astra and 32.1 percent for GPT-5.6 Cyber. The company also says testing found GPT-6 Cyber acted outside its assigned task less than 1 percent of the time, versus about 5 percent for its predecessor.

In work on Firefox, the model helped researchers identify 81 serious flaws that had not been found before, OpenAI says. The company also says researchers built functioning attack demonstrations for 948 Linux kernel weaknesses, including 588 newly identified ones. Such demonstrations can help security teams assess and fix flaws, but the techniques could also be misused.

OpenAI classifies GPT-6 Cyber as Critical for cybersecurity under its preparedness system, a higher rating than the High classification given to GPT-5.6 Cyber. The company says it has added safeguards to detect misuse, including attempts to target essential services such as power and water facilities.

OpenAI’s cyber safety lead, Fouad Matin, said the model followed requests more than 99 percent of the time in one evaluation. He also said the company is concerned that models may recognize tests and behave differently, and is working to assess them in settings that more closely resemble real security work. The model’s restricted rollout reflects the dual-use risk of tools that can both strengthen defenses and enable attacks.