OpenAI Cyber at the Frontier Keynote: Daybreak Platform and GPT-6 Cyber Benchmarks (46:05)

Opening: Greg welcomes the audience, saying change in cybersecurity is accelerating and this is a moment for collective action.

A chart shows progress of the latest cyber model, to be released soon to trusted defenders, a step-function improvement in performance.

Introducing the Daybreak platform.

GPT-6 Cyber was released last week, described as the most capable and most aligned model to date.

ExploitGym benchmark: GPT-6 Cyber scores 52%, Anthropic's Mythos 5.2 about 38%.

PatchBench evaluation: the upcoming cyber model scores around 80%; the challenge is patching as fast as vulnerabilities are found.