OpenAI expands Daybreak with specialist cyber model

OpenAI has expanded its Daybreak cybersecurity access programme with two new tiers and a specialist model, GPT-5.6-Cyber, for approved defensive security users.

Daybreak Blue gives authorised users access to OpenAI's general-purpose models, including GPT-5.6 Sol, with safeguards adjusted for defensive cyber work. Daybreak Red provides models for more advanced tasks such as vulnerability research, exploit validation and security testing, including the new GPT-5.6-Cyber.

The move reflects a broader effort to widen access to models that would otherwise refuse some security-related requests. Standard protections on mainstream systems can block legitimate work by defenders, particularly in malware analysis, incident response, patch validation and security assessments.

GPT-5.6-Cyber is built on GPT-5.6 Sol and trained for specialised cyber tasks, including zero-day vulnerability discovery and exploit-chain development. It also lowers refusal rates for some higher-risk dual-use requests from approved users in the Daybreak Red tier.

In an internal test measuring how often a model completes advanced cyber requests, GPT-5.6-Cyber completed 95.0% of prompts involving areas such as authentication bypass, privilege escalation and exploit-chain development. That compared with 1.5% for GPT-5.6 Sol and 2.0% for GPT-5.6 Sol when used through Daybreak Blue, according to OpenAI. The earlier GPT-5.5-Cyber model completed 57.3% of the same requests.

Benchmark results

OpenAI said the new model outperformed GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, a benchmark that tests whether an agent can turn known vulnerabilities into working exploits in controlled environments. It also said GPT-5.6-Cyber did better than GPT-5.6 Sol on an internal benchmark focused on identifying novel zero-day flaws and producing proof-of-concept exploits with technical write-ups.

Results were more mixed on another internal test for vulnerability discovery and report writing. Both GPT-5.6 Sol and GPT-5.6-Cyber improved on GPT-5.5-Cyber, but GPT-5.6-Cyber trailed GPT-5.6 Sol because it sometimes produced shorter, less detailed reports.

On ExploitBench, which tests an agent's ability to turn a V8 vulnerability into a full exploit under tighter constraints, GPT-5.6 Sol performed best in the standard 300-turn setting. OpenAI said the gap narrowed when the limit was extended to 600 turns.

One external user, SpecterOps, said the specialist model had improved vulnerability research.

"GPT-5.6 Cyber is materially improving our specialist vulnerability-research workflows: it reasons more accurately about real exploit constraints, tracks complex state better, and has completed work in under a day that earlier models had not resolved after weeks of intermittent effort. In a governed Trusted Access environment, reducing unnecessary refusals helps authorized researchers preserve momentum and spend more time validating findings and turning them into defensive value," said Jared Atkinson, Chief Technology Officer, SpecterOps.

Vulnerability work

OpenAI said it has already used GPT-5.6-Cyber in research on real software projects. In one case, the model helped uncover two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome, that could be chained to corrupt memory and escape the V8 heap sandbox.

Researchers validated the findings and reported them to Google through coordinated disclosure. Google fixed the issue and assigned it CVE-2026-15903, which OpenAI described as a high-severity V8 vulnerability involving an omitted safety check during integer conversion that could allow memory access outside expected bounds and potentially arbitrary code execution inside Chrome's sandbox.

OpenAI also said GPT-5.6-Cyber had identified at least five vulnerabilities in a widely used mobile operating system, three critical flaws in a popular database, and more than 400 vulnerabilities linked to privilege escalation in a well-known operating system kernel. It said it is working with partners and open-source maintainers to disclose and fix those issues.

Access controls

OpenAI said both Daybreak Blue and Daybreak Red are limited to approved individuals and organisations carrying out authorised work. Access controls include identity verification, account security measures, monitoring, approved-use restrictions and legal attestations.

The company is also tightening security around the programme. Daybreak users of Codex are being urged to move from full-access mode to auto-review mode, which checks actions requiring elevated permissions before execution and can block requests judged to carry a significant risk of destructive behaviour.

In addition, all individual Daybreak accounts will be required to use hardware security keys. OpenAI also said it is working on further monitoring and other security measures while continuing alignment testing for future Daybreak releases.

Under its Preparedness Framework, OpenAI said both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as reaching a High threshold for cybersecurity capability while remaining below the company's Critical threshold.