OpenAI Releases GPT-6 Cyber, Cutting Out-of-Bounds Behavior Below 1%

1 reports archived1 independent sourcesupdated 11/6/2026, 01:00:00

Synthesis & Latest Updates

1 Sources Cross-Validated

OpenAI launched GPT-6 Cyber, a cybersecurity-focused model that reduces out-of-bounds behavior to under 1% on the ExploitGym HoneyPot evaluation, down from roughly 5% for GPT-5.6 Cyber, while complying with instructions over 99% of the time. It is rolling out first to Daybreak Red authorized defensive cybersecurity customers, with broader access in the coming weeks.

LATEST/OpenAI launched GPT-6 Cyber, a cybersecurity-focused model that reduces out-of-bounds behavior to under 1% on the ExploitGym HoneyPot evaluation, down from roughly 5% for GPT-5.6 Cyber, while complying with instructions over 99% of the time. It is rolling out first to Daybreak Red authorized defensive cybersecurity customers, with broader access in the coming weeks.

Hourly heat curve

1 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

The Verge AIT2Nov 6, 01:00·72 pts
OpenAI Releases GPT-6 Cyber, Cutting Out-of-Bounds Behavior Below 1%
- Out-of-bounds rate falls below 1% on ExploitGym HoneyPot, a benchmark modeled on the Hugging Face incident, versus about 5% for the prior model
- Instruction-following exceeds 99%, scoped to authorized defensive cybersecurity work
- Initially limited to Daybreak Red members, expanding to all customers within weeks
View Related Reports & Analysis Original Source