What it is
OpenAI’s restricted Daybreak cybersecurity program is being evaluated through model benchmarks, product releases and integrations for cyber task performance.
Overview
OpenAI’s GPT-6 Sol is available only through its restricted Daybreak cybersecurity program, with earlier tests showing Daybreak ahead of a cheaper DeepSeek rival. OpenAI introduced Daybreak Red Team and Daybreak Defense Factory to identify and fix software vulnerabilities, with Semgrep compatibility. OpenAI also showed a pre-release cyber system scoring 77% on ExploitGym, while Tencent’s Hy4 posted strong CWE-Bench-AA results.
How it could play out
Why it matters
The evaluation matters for AI labs, cyber defenders, regulators and enterprises weighing restricted frontier cyber tools against cheaper open-weight alternatives.
Key details
• Daybreak completed 80.8% of cyber objectives in cited Vals AI testing. • DeepSeek V4.1 Flash 1017 completed 72.8% and cost under $3 per run. • Daybreak cost about $22 per run in the cited testing. • The internal OpenAI cyber model scored 77% on ExploitGym. • Tencent Hy4 scored 76% on CWE-Bench-AA and 58% on the Artificial Analysis Cyber Index.
Tilt can alert you when its scenarios shift, or send you a recommendation when one crosses a threshold.
Stories in this saga
Portfolio
Portfolio
How it could play out
Saga, how it could play out
SagaOpenAI Daybreak cybersecurity model validation
OpenAI treats the 77% ExploitGym result as proof that advanced cyber agents are nearing sensitive operational thresholds. The most likely path is a controlled rollout limited to vetted defensive partners, with stronger monitoring, red-team review, and internal-use restrictions.
The benchmark result gives regulators a concrete data point for classifying advanced AI cyber systems. U.S. agencies and the EU AI Office are likely to push for reporting, evaluation, and access controls before these models are widely deployed.
The 77% score intensifies competition among frontier AI labs and national AI programs. Rival developers try to demonstrate comparable or better cyber-agent performance, while enterprises increase budgets for AI-assisted vulnerability management.
A smaller but important possibility is that the headline is followed by evidence that AI cyber agents are already being used in real exploit chains or near-miss incidents. Even if OpenAI is not directly involved, the market would reassess software security risk and the urgency of defensive AI adoption.
Portfolio
Portfolio