# OpenAI Splits Daybreak Into Two Tiers, Unveils GPT-5.6-Cyber for Defenders > Source: > Published: 2026-08-11 03:35:16+00:00 OpenAI just changed how cybersecurity professionals get access to its most capable models, and the timing is not accidental. On August 10, 2026, the company announced it is dividing its Daybreak program into two distinct tiers, Blue and Red, while introducing a new purpose-built model called GPT-5.6-Cyber that answers far more security-related requests than its general-purpose counterpart. For the defenders who rely on frontier AI to find vulnerabilities before attackers do, the change addresses a complaint that has followed [OpenAI](https://www.kobaran.com/tag/OpenAI) since Daybreak launched earlier this year: standard safety filters were blocking legitimate security work along with malicious requests. OpenAI says the new structure is meant to fix that imbalance without opening the door to abuse. The announcement lands at a sensitive moment. It comes just days after reports that OpenAI’s unreleased Astra model may cross the “Critical” cybersecurity risk threshold under the company’s own Preparedness Framework, and weeks after OpenAI disclosed that one of its own AI systems was involved in an unauthorized intrusion into Hugging Face. OpenAI is using the Daybreak expansion to draw a firm line between that incident and its newest cybersecurity tools. ## What Daybreak Blue and Daybreak Red Actually Offer OpenAI built the two tiers to match different levels of risk and different kinds of security work. ### Daybreak Blue: broader access, fewer restrictions Daybreak Blue gives verified defenders access to OpenAI’s general-purpose frontier models, including GPT-5.6 Sol, but with the system-level cybersecurity filters loosened for approved users. OpenAI positions this as the recommended entry point for most security teams, covering tasks like vulnerability discovery, secure code review, malware analysis, incident response, and patch validation. ### Daybreak Red: specialized access for higher-risk work Daybreak Red is reserved for a smaller, more heavily vetted group of users doing vulnerability research, exploit validation, and offensive security testing. This is the tier that unlocks GPT-5.6-Cyber, a version of GPT-5.6 Sol that OpenAI trained specifically to be more useful, and less likely to refuse, for advanced dual-use cybersecurity tasks such as identifying zero-day vulnerabilities and building exploit chains. | Tier | Model Available | Primary Use Cases | Access Level | |---|---|---|---| | Daybreak Blue | GPT-5.6 Sol (guardrails adjusted) | Vulnerability discovery, code review, malware analysis, incident response, patch validation | Broad, for verified defenders | | Daybreak Red | GPT-5.6-Cyber | Exploit-chain development, zero-day research, penetration testing, security validation | Narrow, tighter vetting required | ## Why OpenAI Built a Separate Model Instead of Just Loosening Filters According to OpenAI, even Daybreak Blue access still triggers refusals on a subset of prompts that are considered highly dual-use, such as penetration testing on live production systems. GPT-5.6-Cyber was trained to close that remaining gap for authorized Red-tier users. The company measured the difference using an internal benchmark it calls the Advanced Cybersecurity Completion Rate, which tracks how often a model will respond to requests involving exploit-chain development, authentication bypass, and privilege escalation. OpenAI reports that GPT-5.6-Cyber completed 95 percent of these requests, compared with just 1.5 percent for GPT-5.6 Sol under its default safeguards and 2 percent for the same model accessed through Daybreak Blue. The predecessor model, GPT-5.5-Cyber, completed 57.3 percent of similar requests, a rate OpenAI says drew persistent complaints from security researchers who found it too restrictive for real investigative work. ### Capability results are more mixed than the completion-rate numbers suggest OpenAI’s own capability testing shows GPT-5.6-Cyber does not lead across the board. On ExploitGym, a benchmark that tests whether an AI agent can turn a known vulnerability into a working exploit achieving code execution, GPT-5.6-Cyber outperformed both GPT-5.6 Sol and GPT-5.5-Cyber. But on the company’s internal vulnerability discovery and report-writing evaluation, GPT-5.6-Cyber scored below general-purpose GPT-5.6 Sol, which OpenAI attributes to the specialized model producing shorter, less detailed reports. On ExploitBench, a harder exploitation task run inside the V8 JavaScript sandbox with limited information provided to the agent, general-purpose GPT-5.6 Sol actually performed best within a standard 300-turn limit, though the gap narrowed when the agent was given more turns to work with. It is worth noting that all of these figures come from OpenAI’s own internal testing and have not yet been independently replicated by outside evaluators, a point several outlets covering the announcement have flagged. ## A Real-World Vulnerability Find Backs Up the Marketing The most concrete evidence OpenAI offered was not a benchmark score but an actual security find. The company says it used GPT-5.6-Cyber to investigate V8, the JavaScript engine that powers Chrome, and the model surfaced two previously unknown vulnerabilities that could be chained together to corrupt memory and escape the V8 heap sandbox. One of those flaws, a compiler bug where a skipped safety check could let an attacker read or write memory inside Chrome’s sandbox, was reported to Google through coordinated disclosure, patched, and assigned CVE-2026-15903. OpenAI also says its Daybreak partners and open-source maintainers are working through disclosure on several other findings, including at least five vulnerabilities in a widely used mobile operating system, three critical flaws in a popular database with a remote path to code execution, and more than 400 privilege-escalation issues in a widely deployed operating system kernel. OpenAI did not name the specific software involved in those additional cases. SpecterOps, a security firm that tested the model early, offered one of the clearest independent reactions to date. The firm’s CTO, Jared Atkinson, said the model completed work in under a day that earlier models had failed to resolve after weeks of intermittent effort, according to OpenAI’s announcement. ## How OpenAI Is Governing Access to the New Tools Getting into either tier is not automatic. OpenAI requires identity verification, account security checks, ongoing monitoring, approved-use restrictions, and legal attestations, with separate application tracks for individuals and organizations. The company is also tightening how its Codex coding agent operates for Daybreak customers, encouraging a shift from full-access mode to an auto-review mode that checks actions requiring elevated permissions before they execute. Starting September 1, 2026, hardware security keys will be mandatory on every individual Daybreak account. OpenAI says a fuller system card covering additional evaluations of GPT-5.6-Cyber will follow at a later date. ### Where GPT-5.6-Cyber sits on OpenAI’s own risk scale Under OpenAI’s Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were rated High for cybersecurity capability, one step below the framework’s Critical threshold. That distinction matters because it arrived days after reports that OpenAI’s upcoming Astra model may cross into Critical territory. OpenAI used the Daybreak announcement to reiterate that GPT-5.6-Cyber played no role in the unauthorized access to Hugging Face’s systems, and that no model connected to that incident is scheduled for release. ## How Daybreak Has Evolved Since Its Launch The two-tier structure builds on a program OpenAI introduced in stages over the past few months rather than all at once. OpenAI launched the full version of GPT-5.5-Cyber on June 22, 2026, alongside a Daybreak Cyber Partner Program that included Accenture, CrowdStrike, Cisco, IBM, and Palo Alto Networks, along with Patch the Planet, an open-source remediation initiative launched jointly with Trail of Bits. At that time, OpenAI reported GPT-5.5-Cyber scoring 85.6 percent on the CyberGym benchmark against 81.8 percent for standard GPT-5.5, and 39.5 percent versus 25.95 percent on ExploitGym. More than 30 open-source projects signed on to participate, and an initial five-day sprint reportedly surfaced hundreds of issues with dozens of patches merged. Daybreak Blue and Red effectively replace what had been a single Trusted Access track with a two-rung system: a loosened general-purpose model for the broad base of verified defenders, and a specialized, refusal-light model reserved for the smaller group of researchers doing exploit development and red-team work. ### What this means going forward For security teams already inside the Daybreak ecosystem, the practical shift is faster, less-filtered access to tools they say they already needed. For everyone else, the announcement is another data point in a fast-moving debate over how much offensive capability AI labs should put directly into the hands of vetted outsiders, even when the stated goal is defense. With OpenAI’s own Astra model reportedly approaching a higher risk tier and additional evaluations of GPT-5.6-Cyber still to come, this is unlikely to be the last update to the Daybreak program.