Frontier AI for Defenders Too! OpenAI Splits Daybreak Into Two Tiers With New Model GPT-5.6-Cyber

Hey everyone, it's me, Shiichan! Today I've got news that's about defense, not offense. OpenAI is putting a serious cybersecurity model into the hands of vetted defenders, and reading through it made me sit up a little straighter.

OpenAI News openai.com

What was announced?

OpenAI's News page announced it is expanding its cybersecurity program, Daybreak, and introducing a new model called GPT-5.6-Cyber. The goal is to put frontier-level AI in the hands of legitimate defense teams before attackers can automate AI-driven attacks at scale.

Daybreak now offers two access tiers.

• Daybreak Blue: gives access to the general-purpose GPT-5.6 Sol with safeguards tailored for defensive work. It supports vulnerability discovery, secure code review, malware analysis, incident response, and patch validation, and it's the recommended starting point for most defenders.
• Daybreak Red: gives access to a purpose-trained cybersecurity model, GPT-5.6-Cyber, for authorized vulnerability research, exploit validation, and security testing.

GPT-5.6-Cyber is built on GPT-5.6 Sol and trained to improve performance on specialized tasks like finding zero-day vulnerabilities and developing exploit chains, while reducing the refusals that used to get in the way of legitimate defensive work.

The story so far

GPT-5.6 Sol normally screens cybersecurity-related requests with system-level safeguards in production. That prevents misuse, but it can also block legitimate defensive work. Daybreak Blue removes those guardrails, but GPT-5.6 Sol would still refuse certain highly dual-use prompts, like pentesting production systems.

On top of that, the previous model, GPT-5.5-Cyber, had a refusal problem that security researchers kept flagging. On OpenAI's internal Advanced Cybersecurity Completion Rate evaluation, which measures how often a model responds to requests involving exploit-chain development, authentication bypass, privilege escalation, and similar advanced scenarios, GPT-5.5-Cyber completed only 57.3% of them.

What changes

On that same evaluation, GPT-5.6-Cyber through Daybreak Red completes 95.0% of requests. Compare that to 1.5% for GPT-5.6 Sol with standard guardrails and 2.0% even through Daybreak Blue, and you can see how much a purpose-built model changes the picture.

For authorized defense teams and security researchers, that means:

• Going from vulnerability discovery to building an exploit chain without getting stuck on refusals along the way
• Sustaining long investigations across large, unfamiliar codebases, forming and testing hypotheses with the model alongside you
• SpecterOps CTO Jared Atkinson said work that used to take weeks of intermittent effort with earlier models was finished in under a day

Access is still limited to authorized work only, not something anyone can just pick up, so it's worth reading this alongside the safety measures below.

Dive Deep

Here's how it performs on benchmarks.

• ExploitGym, which tests whether an agent can turn a known vulnerability into a working exploit: GPT-5.6-Cyber outperforms both GPT-5.6 Sol and GPT-5.5-Cyber.
• An internal zero-day discovery evaluation, where a model is handed an open-source repo and asked to produce a proof-of-concept exploit plus a technical write-up: GPT-5.6-Cyber (Daybreak Red) outperforms GPT-5.6 Sol (Daybreak Blue).
• An internal vulnerability discovery and report-writing evaluation: here GPT-5.6 Sol actually scores higher than GPT-5.6-Cyber. OpenAI attributes this to GPT-5.6-Cyber tending to produce shorter, less detailed reports.
• ExploitBench, which tests turning a V8 vulnerability into a full exploit with defenses like the V8 sandbox still enabled and less information to work with: in the standard 300-turn setting, GPT-5.6 Sol is more token-efficient and performs best; the gap narrows when the turn limit is extended to 600.

After training finished, OpenAI also put GPT-5.6-Cyber to work on real software.

• It found two previously unknown vulnerabilities in Chrome's V8 engine that could lead to memory corruption. OpenAI reported them to Google through coordinated disclosure, Google fixed them, and the issue is now tracked as CVE-2026-15903.
• At least five vulnerabilities in a popular mobile OS, including a chain from an untrusted app to local privilege escalation.
• Three critical vulnerabilities in a popular database, including a path to remote code execution.
• Over 400 vulnerabilities leading to privilege escalation in a popular OS kernel.

OpenAI says it's working with Daybreak partners and the open-source community to disclose and fix these.

Under OpenAI's Preparedness Framework, GPT-5.6 Sol was rated High for cybersecurity capability, below the Critical threshold. GPT-5.6-Cyber was evaluated the same way and also lands at High, not Critical. OpenAI also noted that GPT-5.6-Cyber was not involved in the earlier Hugging Face incident.

Running a model with reduced safeguards carries more risk than standard usage, so OpenAI is pairing the release with several safety measures:

• Daybreak Blue and Red access is limited to approved individuals and organizations that go through identity verification, account security checks, monitoring, usage restrictions, and legal attestations.
• Daybreak users on Codex are encouraged to switch from full-access mode to auto-review mode, which checks actions that need elevated permissions before they run.
• All individual accounts will be required to use hardware security keys starting September 1, 2026.
• OpenAI says it's also working on additional measures like improved monitoring, plus prioritizing alignment training and testing for future Daybreak releases.

Recommended best practices include running security workflows in sandboxed environments isolated from production systems and the open internet, monitoring agent actions, and clearly scoping which systems and actions are authorized.

You can apply for the program at the Daybreak partners page. OpenAI recommends starting with Daybreak Blue, and teams doing advanced vulnerability research, exploit development, or red teaming can apply for Daybreak Red.

Wrap-up

• Daybreak now has two tiers: Daybreak Blue (GPT-5.6 Sol plus defense-tailored safeguards) and Daybreak Red (the purpose-built GPT-5.6-Cyber model).
• GPT-5.6-Cyber completes 95.0% of advanced cyber task requests on OpenAI's internal benchmark, versus 1.5% for GPT-5.6 Sol and 57.3% for the earlier GPT-5.5-Cyber.
• Real-world use already found a Chrome V8 vulnerability (CVE-2026-15903), plus vulnerabilities in a mobile OS, a database, and an OS kernel.
• Under OpenAI's Preparedness Framework, GPT-5.6-Cyber is rated High for cybersecurity capability, below the Critical threshold.
• Access comes with real guardrails: identity verification, legal attestations, and mandatory hardware security keys.
• A solid step forward for authorized security researchers, red teams, and enterprise security teams who want to move defensive work faster.