OpenAI has expanded its Daybreak cybersecurity program, introducing a specialized model called GPT-5.6-Cyber alongside two new access tiers designed for professional security researchers. The move, announced on August 10, 2026, gives vetted defenders a far more capable tool for finding and analyzing software vulnerabilities — arriving in the same news cycle that saw the company pause its next-generation Astra model over cyber-risk concerns.

The expansion underscores how breaking AI news is increasingly shaped by the tension between offensive and defensive cyber capabilities. OpenAI is betting that putting powerful, lower-refusal models into trusted hands can outpace the threat of AI-driven attacks.

What is the Daybreak program?

Daybreak launched in May 2026 as a way for cybersecurity professionals to use OpenAI's models for vulnerability research. Because OpenAI's commercially available large language models include filters that block most cybersecurity prompts, Daybreak removes many of those restrictions so participants can automate time-consuming security tasks such as scanning code, verifying patches, and analyzing malware.

With the August update, OpenAI split access into two tiers:

  • Daybreak Blue — described by the company as the "starting point for most defenders." Software teams can use it to scan their own code for vulnerabilities, confirm whether patches actually work, and analyze external code such as malicious files.
  • Daybreak Red — intended for more advanced work, such as determining whether a newly discovered vulnerability can actually be exploited. This tier is powered by GPT-5.6-Cyber.

GPT-5.6-Cyber: a model trained to refuse less

The centerpiece of the update is GPT-5.6-Cyber, a custom version of OpenAI's flagship GPT-5.6 Sol algorithm that has been optimized for advanced cybersecurity tasks such as hunting for zero-day vulnerabilities.

To measure the difference, OpenAI built a benchmark tracking how often a model refuses advanced cybersecurity requests. The results illustrate a sharp jump in capability:

  • The broadly available GPT-5.6 Sol completed just 1.5% of the benchmark's requests.
  • Daybreak Blue answered 2% of prompts.
  • GPT-5.6-Cyber completed 95%.

On a separate benchmark called ExploitGym2, which evaluates exploit-development capabilities, GPT-5.6-Cyber scored slightly higher than the standard GPT-5.6 Sol model, according to OpenAI.

Importantly, the model is not a free-for-all. OpenAI said GPT-5.6-Cyber was trained during development to block particularly risky requests — for example, customers cannot use it to probe weak points in live production systems.

Real-world bug discoveries

OpenAI says its own engineers have already put GPT-5.6-Cyber to work in the field. According to the company, the model helped uncover two high-severity vulnerabilities in Google Chrome that allowed attackers to overwrite parts of the browser's memory with malicious code before Google issued patches. In a separate internal project, OpenAI reported finding three critical vulnerabilities in an unnamed but widely used database.

These discoveries are meant to demonstrate that the model can deliver defensive value — finding flaws so they can be fixed — rather than simply enabling attacks.

Tighter guardrails alongside looser models

In a notable juxtaposition reported by outlets including Axios, TechCrunch, and VentureBeat, the Daybreak expansion arrived the same day that coverage of OpenAI's decision to pause its Astra model over cybersecurity concerns was still fresh. As The Next Web noted, OpenAI had paused a model over cyber risk on Friday, then shipped one trained to refuse less on Monday.

OpenAI is addressing that risk through access controls rather than blanket refusals. Starting next month, Daybreak participants will be required to log in with hardware security keys, and the company is rolling out new monitoring features designed to detect unauthorized or abusive activity within the program.

Why it matters

The Daybreak expansion reflects a broader industry shift. As AI-led attacks multiply, defenders argue they need AI tools that are at least as capable as those available to attackers — and that traditional safety filters, which block nearly all cybersecurity use cases, leave legitimate researchers hamstrung.

OpenAI's approach attempts to thread that needle: rather than loosening restrictions for everyone, it concentrates powerful capabilities inside a gated, monitored program with verified participants. Whether that balance holds as the model reaches more hands will be one of the defining questions for the latest AI developments in the months ahead.

For now, the message from OpenAI is that the defensive window is narrowing — and that the best way to close it is to give the right people a sharper tool.

Stay Ahead of AI

Cybersecurity is just one front in a fast-moving field. For continuous coverage of model releases, safety debates, and industry shifts, follow our AI industry coverage and read more AI news →