OpenAI hands vetted security firms a model that answers 95% of hacking prompts

On Monday, OpenAI split its Daybreak cyber-defense program into Blue and Red tiers and delivered GPT-5.6-Cyber. The company passed the hacking model to a handful of vetted security firms that are trying to beat AI attacks.

Earlier this year, OpenAI launched Daybreak to give unreleased frontier models to private organizations and governments for defensive work.

The program now splits customers into two, Blue and Red. Blue is powered by GPT-5.6 Sol, the company’s general purpose frontier model, with production guardrails that normally filter security prompts removed.

Daybreak splits into Blue and Red lanes

OpenAI calls it the “recommended starting point for most defenders.” The Blue tier performs vulnerability discovery, secure code review, malware analysis, incident response, and patch validation.

Red provides access to purpose-built cybersecurity models for authorized vulnerability research, exploit validation, and security testing. This is the only tier where GPT-5.6-Cyber exists.

OpenAI said Red customers will be closely monitored and supervised for usage because the model is much more capable of malicious cyber tasks than Sol.

GPT-5.6-Cyber is built on GPT-5.6 Sol but trained to find zero-day vulnerabilities, build chains of exploits, and deny fewer high-risk, dual-use requests.

OpenAI measured the gap with an internal benchmark dubbed the Advanced Cybersecurity Completion Rate. Each model is tested for exploit-chain development, authentication bypass, and privilege escalation.

GPT-5.6-Cyber managed to answer 95.0% of the questions. GPT-5.6 Sol managed to answer 1.5% and 2.0% via Daybreak Blue. 57.3% for older GPT-5.5-Cyber. That number was tied to complaints from researchers who kept getting rejected, according to OpenAI.

OpenAI loosens guardrails for 16 vetted vendors

OpenAI has also rolled out a partner program with 16 cybersecurity providers. Organizations access the frontier models via the security services they already buy. Partners include IBM, CrowdStrike, Accenture, Ernst & Young, KPMG, Palo Alto Networks, Cisco, Cloudflare, and Sophos.

GPT-5.6-Cyber is only available to “trusted customer partners.” They include Accenture, IBM, CrowdStrike, and Cloudflare.

The model “has completed work in under a day that earlier models had not resolved after weeks of intermittent effort,” SpecterOps CTO Jared Atkinson said. He said that by eliminating unnecessary rejections, authorized researchers can spend more time on validating findings. OpenAI said a more complete system card for GPT-5.6-Cyber will come out later.

“Models running with reduced safeguards carry risks beyond standard model usage, whether from misuse or misalignment,” the company wrote. It still believes that “democratizing access to frontier intelligence for defenders is crucial.”

In 122 test runs, agents from OpenAI and Anthropic broke rules 19 times, the UK AI Security Institute said on August 5. Two traced to GPT-5.6-Sol. The rest mostly went to Anthropic’s Mythos 5.

Daybreak arrived after Anthropic shipped Mythos, its own cyber-focused model. OpenAI said it is slowing work on a separate model, Astra, to build better controls.

Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.