GPT-6 Cyber cut out-of-bounds actions below 1% in tests
OpenAI says the model is better at finding software weaknesses, and that in its tests it went out of bounds less than 1% of the time, down from about 5% for its predecessor.
OpenAI released GPT-6 Cyber to select Daybreak Red customers for authorised security testing.
OpenAI says the model is better at finding vulnerabilities and less likely to go out of bounds.
In OpenAI's testing, out-of-bounds activity fell to under 1% of the time, from around 5% with GPT-5.6 Cyber.
Reports on a Linux kernel study say the model was used to develop working exploits for 948 flaws.
An earlier OpenAI model helped a researcher find a severe Ubuntu-related bug; there is no known evidence it was exploited.