OpenAI is expanding its Daybreak initiative to offer GPT-5.6-Cyber for authorised defensive security work.
The programme gives approved defenders two access routes: Daybreak Blue for frontier general-purpose models, and Daybreak Red for purpose-trained cybersecurity models. OpenAI positions Blue as the entry point for most security teams. Red targets authorised vulnerability research, exploit validation, and security testing.
Harpreet Sidhu, Global Lead for Accenture Cybersecurity, said: “Cybersecurity teams are under growing pressure to not just find vulnerabilities quickly, but to now fix them quickly. This change requires partners who can turn vulnerability insights into immediate action across complex systems.”
Daybreak separates standard defence work from advanced research
Daybreak Blue provides GPT-5.6 Sol with system-level cybersecurity guardrails removed. OpenAI says those standard protections can block legitimate activity during incident detection, investigation, vulnerability management, security assessment, secure code review, malware analysis, and patch validation.
Some boundaries remain. GPT-5.6 Sol will still refuse highly dual-use requests through Daybreak Blue, including prompts involving production-system penetration testing.
Daybreak Red introduces GPT-5.6-Cyber, a specialist model built on GPT-5.6 Sol. OpenAI trained it for advanced tasks, including zero-day discovery and exploit-chain development, and it reduces refusals for certain higher-risk requests. That makes Red suitable only where an organisation can document authorisation, constrain system access, and review the model’s actions.
OpenAI’s internal Advanced Cybersecurity Completion Rate puts GPT-5.6-Cyber at 95 percent for requests involving advanced cyber scenarios. GPT-5.6 Sol completed 1.5 percent with standard safeguards, according to the company. Daybreak Blue raised that figure to 2 percent, and GPT-5.5-Cyber reached 57.3 percent.
Those figures measure willingness to respond. They do not measure a model’s accuracy, the safety of an output, or the quality of a remediation.
Benchmark gains come from isolated environments
OpenAI reports that GPT-5.6-Cyber outperformed GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, an evaluation that tests whether an agent can turn known vulnerabilities into functioning exploits capable of arbitrary code execution. Its testing used security-hardened, isolated environments with monitoring for misaligned behaviour.
A different result appeared on ExploitBench. GPT-5.6 Sol with Daybreak Blue solved V8 exploitation tasks most efficiently in the standard 300-turn setting. OpenAI says the gap between the models narrowed after it allowed 600 turns.
No direct production comparison exists in these tests. ExploitGym and ExploitBench ran inside controlled environments, where OpenAI set the target systems, constraints, and agent limits. Enterprise security teams would need to test model behaviour against their own code repositories, tool permissions, incident procedures, and approval paths before allowing autonomous execution.
GPT-5.6-Cyber also performed below GPT-5.6 Sol in OpenAI’s internal Vulnerability Discovery and Report Writing evaluation. The company attributes that result to shorter, less detailed vulnerability reports from the specialist model. That trade-off matters for teams whose vulnerability-management process depends on reports that developers, risk owners, and incident staff can act on without reworking the findings.
OpenAI cites V8 findings and partner use
OpenAI says it used GPT-5.6-Cyber to investigate V8, Chrome’s JavaScript engine, and found two previously unknown vulnerabilities that attackers could chain to corrupt memory and escape V8’s heap sandbox. OpenAI researchers validated the findings, reported them to Google through coordinated disclosure, and Google assigned CVE-2026-15903 after fixing the issue.
The reported flaw involved V8’s optimising compiler. OpenAI says the compiler skipped a safety check during an integer conversion, allowing an undefined value to produce an unexpectedly large number. When code used that value as an array index, the compiler could omit a bounds check and permit memory reads or writes outside the intended object.
OpenAI also says GPT-5.6-Cyber identified at least five vulnerabilities in a popular mobile OS. The company says one chain ran from an untrusted application to local privilege escalation. It reports three high-severity flaws in a popular database, including a remote route to code execution. OpenAI also reports more than 400 privilege-escalation vulnerabilities in a popular OS kernel.
The company has not named the mobile operating system, database, or kernel in the post. It says it is working with Daybreak partners and open-source community members on disclosure and remediation.
SpecterOps CTO Jared Atkinson provided the most concrete customer account in OpenAI’s announcement.
“[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,” Atkinson said.
Atkinson also described a governed Trusted Access environment. He said reduced refusals allow authorised researchers to spend more time validating findings and converting them into defensive work.
Access controls define the operational model
OpenAI says Daybreak requires identity verification, account-security controls, monitoring, approved-use restrictions, and legal attestations. It will require hardware security keys for all individual Daybreak accounts from 1 September 2026.
Codex users receive another control point. OpenAI encourages Daybreak customers to switch from full-access mode to auto-review mode through defaults and interface features. Auto-review checks actions that require elevated permissions before execution, and it can block requests that pose a risk of destructive behaviour.
The company recommends that organisations run security workflows in controlled environments without access to sensitive production systems or the open internet. Teams should test sandbox boundaries regularly. OpenAI also recommends that users review tool calls outside the Codex sandbox, apply human oversight to higher-risk workflows, and define authorised systems and actions through scoped permission profiles.
Managed-service providers and security vendors feature prominently in the supporting material. Sophos says it applies Daybreak models across its MDR, incident-response, and advisory services.
Cloudflare CTO Dane Knecht said the company intends to pair Daybreak capabilities with Cloudflare’s global network to help security teams turn threat intelligence into action. “AI presents huge potential in cybersecurity, but its true power lies in how practitioners apply it in the real world,” he said.
OpenAI recommends Daybreak Blue for most defenders. Daybreak Red requires a more demanding operating model, because its value comes from responding to work that standard systems would decline. For enterprises, that places sandboxing, scoped permissions, hardware-backed account protection, action review, and documented authorisation ahead of any broad rollout of GPT-5.6-Cyber.