OpenAI announced on August 10, 2026, that it is expanding its Daybreak cybersecurity program into two access tiers and releasing a new purpose-trained model, GPT-5.6-Cyber. The move gives approved defenders a clearer path to powerful AI tools while keeping the most sensitive capabilities under stricter control. The two tiers, Daybreak Blue and Daybreak Red, replace what had been a single Trusted Access track.
Two Tiers, Two Levels of Access
Daybreak Blue opens frontier general-purpose models, including GPT-5.6 Sol, to approved defenders for everyday security work. OpenAI says system-level safeguards on GPT-5.6 Sol screen cybersecurity-related requests to prevent misuse, but those same safeguards also block legitimate defensive work. Daybreak Blue removes those safeguards for verified users.
Daybreak Red gates the new GPT-5.6-Cyber model behind tighter vetting. This tier is reserved for vulnerability research, exploit validation, and security testing. The two-rung ladder is straightforward: a general-purpose frontier with relaxed guardrails for the broad defender base, and a refusal-light specialist model for the smaller group whose authorized work runs to exploit development and red teaming.
Access to both tiers runs through identity verification, account security requirements, monitoring, approved-use restrictions, and legal attestations. There are separate application paths for individuals and organizations. Beginning September 1, 2026, OpenAI will require hardware security keys on all individual Daybreak accounts.
A Model Built to Say Yes More Often
GPT-5.6-Cyber is a version of GPT-5.6 Sol trained to reduce refusals on advanced cybersecurity tasks. It also improves on specialized work like finding zero-day vulnerabilities and developing exploit chains. OpenAI built an internal evaluation called the Advanced Cybersecurity Completion Rate to measure how often models respond to requests involving exploit-chain development, authentication bypass, privilege escalation, and similar scenarios.
The numbers tell a dramatic story. GPT-5.6-Cyber completes 95.0% of requests on the Advanced Cybersecurity Completion Rate. GPT-5.6 Sol under standard safeguards completes just 1.5% on the same benchmark. Even with Daybreak Blue safeguards removed, GPT-5.6 Sol reaches only 2.0%. The prior model, GPT-5.5-Cyber, completed 57.3%. OpenAI says the refusal rate of GPT-5.5-Cyber generated persistent complaints from security researchers.
The jump in permissiveness is stark, but capability evaluations tell a more qualified story. On ExploitGym, which tests whether agents can turn known vulnerabilities into working exploits achieving code execution in controlled environments, GPT-5.6-Cyber outperforms both GPT-5.6 Sol and GPT-5.5-Cyber. On OpenAI's internal Vulnerability Discovery and Report Writing evaluation, GPT-5.6-Cyber improves over GPT-5.5-Cyber but lands below GPT-5.6 Sol. OpenAI attributes the lower performance to GPT-5.6-Cyber producing shorter, less detailed reports.
On ExploitBench, a harder exploitation task with the V8 sandbox enabled and less information given to the agent, GPT-5.6 Sol performs best within the standard 300-turn limit. The gap narrows when runs extend to 600 turns. All benchmark figures are vendor-reported, and several are on internal benchmarks not replicated by outside evaluators.
Real Vulnerabilities Found
The most significant claims are not benchmarks but vulnerability discoveries. OpenAI used GPT-5.6-Cyber to investigate V8, the JavaScript engine inside Chrome. The model uncovered two previously unknown V8 vulnerabilities that could be chained to corrupt memory and escape the V8 heap sandbox.
The first V8 vulnerability is a compiler bug where a skipped safety check lets an attacker read or overwrite memory inside Chrome's sandbox. OpenAI reported it to Google through coordinated disclosure. Google fixed the issue and assigned it CVE-2026-15903.
The findings extend beyond Chrome. OpenAI lists, without naming affected software, at least five vulnerabilities in a popular mobile operating system, including a privilege-escalation chain from an untrusted app. The company also lists three critical vulnerabilities in a popular database, including a remote path to code execution. And it lists more than 400 privilege-escalation vulnerabilities in a popular operating system kernel. All listed vulnerabilities are moving through disclosure with Daybreak partners and open-source maintainers.
Early testers saw the impact firsthand. Jared Atkinson, CTO of SpecterOps, tested the model early and said it "has completed work in under a day that earlier models had not resolved after weeks of intermittent effort." That kind of work compression is what security teams want, but it also raises the stakes for keeping the model out of the wrong hands.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Governance Tightens
A system card with further evaluations of GPT-5.6-Cyber is planned for a later date. Under OpenAI's Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as High for cybersecurity capability and below the Critical threshold. The assessment lands days after Unite.AI reported that OpenAI's upcoming Astra model may cross the Critical cybersecurity threshold. Jonas Reeve, an AI-generated analyst at Unite.AI, authored that article. Articles authored by Jonas Reeve are AI-generated and reviewed by Unite.AI's editorial team.
OpenAI also reiterated that GPT-5.6-Cyber was not involved in the exploitation of Hugging Face, and no model with that involvement is slated for release. That clarification comes after an earlier incident disclosure related to the Hugging Face platform.
Building on a June Launch
The Daybreak program has been expanding in pieces. On June 22, 2026, OpenAI launched the full version of GPT-5.5-Cyber. That launch included a Daybreak Cyber Partner Program with Accenture, CrowdStrike, Cisco, IBM, and Palo Alto Networks as participants. It also introduced Patch the Planet, an open-source remediation initiative founded with Trail of Bits.
On that earlier release, OpenAI reported GPT-5.5-Cyber reaching 85.6% on CyberGym against 81.8% for GPT-5.5. On ExploitGym, GPT-5.5-Cyber reached 39.5% against 25.95% for GPT-5.5. Per the June 22 post, more than 30 open-source projects committed to participate in Patch the Planet. The initial five-day sprint surfaced hundreds of issues with dozens of patches merged.
The new tiered structure consolidates the program. Daybreak Blue and Red replace the single Trusted Access track. The design reflects the tension OpenAI manages between preventing misuse and enabling legitimate defensive work. The Advanced Cybersecurity Completion Rate shows a dramatic increase in permissiveness for GPT-5.6-Cyber, but the capability evaluations show mixed results across benchmarks.
The vulnerability discoveries carry the most weight. Two previously unknown V8 flaws, a privilege-escalation chain in a mobile OS, critical database issues, and hundreds of kernel vulnerabilities are concrete results. Those findings suggest the model is doing real work, not just passing internal tests.
OpenAI's approach is cautious in structure but aggressive in capability. The two tiers give defenders more tools while keeping the most dangerous capabilities behind stricter vetting. The hardware security key requirement adds another layer. Whether the safeguards hold will depend on how well the vetting works in practice.
The company is also watching its own models closely. The Preparedness Framework assessment keeps both GPT-5.6 Sol and GPT-5.6-Cyber below the Critical threshold. But the Astra report suggests the next model could push past that line. OpenAI has not commented on Astra beyond the existing framework assessments.
For now, the focus is on the two tiers and what they enable. Daybreak Blue serves the broad defender base with relaxed guardrails. Daybreak Red serves the smaller group doing exploit development and red teaming. Both require identity verification, monitoring, and legal attestations. The system card will provide more detail later.
The September 1, 2026 hardware security key requirement applies to all individual Daybreak accounts. That deadline is weeks away. Organizations face separate application paths and their own security requirements.
OpenAI's Daybreak expansion is a bet that giving defenders more capable AI tools will improve security faster than it enables attackers. The benchmark numbers are vendor-reported, and several are internal. The vulnerability discoveries, though, are verifiable through disclosure processes already underway with Google and other partners.
The model found real flaws in real software. That is the measure that matters most.