OpenAI GPT-6 Cyber Reproduces 1724 Linux Kernel Vulnerabilities While Hitting 99.93% Off-Limits Restraint

OpenAI released GPT-6 Cyber under the Daybreak Red program with API and Codex access restricted to approved customers. The model documented 81 high-severity Firefox flaws accepted by Mozilla and submitted 304 critical issues with fixes to kernel maintainers. New monitoring enforces user-permitted actions and blocks industrial control assistance, refined via Dragos lab tests on Daybreak Red models.

Capability gains are paired with alignment claims: without safety prompts the model avoided off-limits targets in 99.93% of cases versus 94.49% for the prior version. Official statements emphasize scope discipline, yet the fivefold increase in kernel exploit output and Critical rating under the Preparedness Framework indicate rapid offensive potential. Independent verification of the 1,724 reproductions remains limited to OpenAI-provided tooling logs.

Dragos noted the dropping skill floor for reconnaissance and vulnerability research. A dedicated critical infrastructure defender group will allow sanctioned testing, while broader Daybreak Red rollout continues. Mozilla patches are already shipping based on the 81 reported flaws.

The pattern shows defensive tooling announcements coinciding with capability jumps that outpace public red-team scrutiny. Contract records and procurement for similar models will determine whether access controls hold once the model reaches more operators.