OpenAI expands Daybreak and announces a model reserved for authorized security activities

OpenAI announced on August 10, 2026 the expansion of its cybersecurity initiative, called Daybreak, and the unveiling of GPT-5.6-Cyber, a model intended for advanced and authorized activities in that domain (Releasebot).

The initiative is organized into two channels, referred to as Daybreak Blue and Daybreak Red, reserved for accredited defenders. According to the documentation reported by the source, Daybreak Blue provides frontier models — including GPT-5.6 Sol — with safeguards calibrated to a broad defensive scope. The planned activities include vulnerability detection, security-oriented code review, malware analysis, incident response, and verification of deployed fixes.

In the same update, the company made multi-agent operation natively available in its programming interface, the Responses API: a primary agent coordinates the work and delegates parts of it to sub-agents. This is a distinct element from the security announcement, but it touches on the same ground, since it allows analysis steps that until now required external orchestration to be automatically chained together.

It should be clearly stated what remains outside the available material. The source does not report the number of already accredited entities, nor the criteria by which accreditation is granted or revoked, nor the operational difference between the two channels beyond their existence. There are, at this stage, no assessments conducted by third parties on the described safeguards, nor independent measures of the models’ capability in the listed activities: the only available description is that of the party offering the service, and should be read as such — the party’s position, not proof.

On this piece, the news currently comes from a single source (Releasebot, which relays updates published by OpenAI); no independent confirmation available. Redazione Zero publishes it for the relevance of its subject matter — selective access to models capable of searching for vulnerabilities in real code is an industrial policy choice, not a product detail — and does not attribute to it the value of a verified fact.

The practical consequence of the announced framework is that the described capabilities are not available in general form: they pass through a filter managed by the company. Those working on defense outside that filter, in public institutions or independent research, depend on the admission criteria, and the source does not publish them. No economic access conditions or opening dates for applications to the two channels are indicated.

Sources: Releasebot — OpenAI updates (aggregator of corporate announcements; primary source: OpenAI communication).