OpenAI Group PBC is expanding a program that enables cybersecurity professionals to use its models for vulnerability research.

Daybreak, as the the program is called, launched in May. OpenAI today added two new access tiers called Daybreak Blue and Daybreak Red. It also announced plans to enhance the guardrails that block unauthorized users from joining the program.

OpenAI’s commercially available large language models include filters that block most cybersecurity prompts. Daybreak removes many of those filters. As a result, program participants can use OpenAI’s LLMs to automate time-consuming vulnerability research tasks. The list of available models includes the company’s flagship GPT-5.6 Sol algorithm.

Daybreak Blue, the first Daybreak access tier that launched today, is intended to be the “starting point for most defenders.” Software teams can use it to scan their code for vulnerabilities and determine whether patches work as expected. Additionally, Daybreak Blue lends itself to analyzing external code such as malware files.

It doesn’t support more advanced tasks such as determining whether a newly discovered vulnerability can be exploited. OpenAI is addressing that use case with Daybreak Red, the second Daybreak access tier it debuted today. It’s powered by a custom version of GPT-5.6 Sol called GPT-5.6-Cyber. 

OpenAI created a benchmark for measuring how often an LLM refuses to perform advanced cybersecurity tasks. The broadly available version of GPT-5.6 Sol completed 1.5% of requests. Daybreak Blue answered 2% of the prompts in the benchmark, while GPT-5.6-Cyber completed 95%.

According to OpenAI, GPT-5.6-Cyber not only supports a broader range of prompts than GPT-5.6 Sol but also performs some tasks better. It compared the LLMs using a benchmark called ExploitGym2 that measures models’ exploit development capabilities. GPT-5.6-Cyber achieved a slightly higher score.

OpenAI developed GPT-5.6-Cyber by optimizing GPT-5.6 Sol to perform advanced cybersecurity tasks such as finding zero-day vulnerabilities. During development, the company also trained the model to block particularly risky requests. For example, customers can’t use the model to find weak points in production systems.

The company says its engineers have used GPT‑5.6‑Cyber to discover two high-severity vulnerabilities in Chrome. The flaws made it possible to overwrite parts of the browser’s memory with malicious code before Google patched them. During another internal project, OpenAI found three critical vulnerabilities in an unnamed popular database. 

The company is rolling out the two new Daybreak access tiers alongside upgrades to the program’s user guardrails. Starting next month, participants will be required to log in with hardware security keys. OpenAI is also implementing new monitoring features designed to detect unauthorized activity.

Image: OpenAI