OpenAI is expanding the scope of Daybreak, the cybersecurity initiative it introduced in May as a counterpart to Anthropic's Project Glasswing, in response to a rapidly evolving AI security landscape where frontier model developers are racing to equip defenders with specialized tools before threat actors can exploit similar capabilities at scale.

The expansion was confirmed in a blog update on the OpenAI website, which revealed that the program will now offer two distinct access tiers: Daybreak Blue and Daybreak Red.

The restructuring comes amid escalating concern following several recent cybersecurity incidents. Models from OpenAI, Anthropic, and Meta all accessed online areas that should have been off-limits from their testing environments. Most recently, OpenAI paused testing on a model called Astra after it reached what the company described as its "critical cybersecurity threshold."

In its blog post, OpenAI acknowledged the growing threat posed by AI-enabled cyberattacks, stating: "The cybersecurity world is rapidly changing -- threat actors will increasingly use AI to conduct cyberattacks at unprecedented speed and scale, including in fully autonomous ways. As these capabilities spread, defenders have a narrowing window to prepare."

Two-Tier Access Structure

The new Daybreak Blue tier, the more basic of the two and the recommended starting point for most defenders, grants users access to frontier general-purpose models including GPT-5.6 Sol. Safeguards on these models have been adjusted to permit authorized defensive security work.

The more advanced Daybreak Red tier provides access to OpenAI's purpose-trained cybersecurity models designed for testing and vulnerability research. Users at this tier will also gain access to GPT-5.6-Cyber, a new model built on GPT-5.6 Sol that has been trained to deliver improved performance across specialized cybersecurity tasks and is less likely to refuse high-risk requests.

According to OpenAI, reducing the model's refusal rate for certain cybersecurity tasks addresses "feedback from security researchers who encountered persistent refusals" from the previous GPT-5.5-Cyber when submitting requests involving advanced cybersecurity scenarios. The company reported that GPT-5.6-Cyber completes 95.0% of requests, a significant increase from the prior model's 57.4%. The improvement reflects a broader industry tension: safety guardrails designed to prevent malicious use can also block legitimate defensive research, pushing labs to develop tiered access models that attempt to serve authorized professionals without exposing dangerous capabilities to the public.

Restricted Access

As has been the case since Daybreak's launch, access remains limited to approved organizations and individuals conducting authorized work, with OpenAI citing the potential risks of misuse. However, the company appears prepared to broaden access, as the blog post includes a link to a page where prospective users can submit an official application to sign up.

The expansion marks the latest effort by a major AI lab to address mounting concerns about AI cybersecurity vulnerabilities, as the industry grapples with the dual challenge of enabling defensive research while preventing malicious exploitation of increasingly capable models.