On 11 August 2026, OpenAI launched GPT-5.6-Cyber, a variant of the GPT-5.6-Sol flagship specifically trained for offensive security tasks. The model is available exclusively through the Daybreak Red programme, a closed access programme that requires identity verification, legal declarations and approved use cases. This is a different approach from previous models: deliberate access restriction as a safety layer.
What can the model do?
GPT-5.6-Cyber is based on GPT-5.6-Sol and further trained on offensive security workflows, such as finding zero-day vulnerabilities and developing exploit chains. OpenAI reports that the model completes 95% of advanced exploit-chain and privilege-escalation prompts, compared with only 1.5% for the standard GPT-5.6-Sol. That is not a marginal improvement, but a structurally different category.
In concrete terms: the model identified vulnerabilities in Chrome's V8 engine and in several unnamed mobile systems and databases. The lower refusal threshold for dual-use security tasks was designed deliberately, and just as deliberately safeguarded.
Access runs through OpenAI's Daybreak programme, which now has two tracks: Daybreak Blue (defensive) and Daybreak Red (offensive). Known partners include Accenture, IBM, Palo Alto Networks and CrowdStrike. Individual access requires hardware security keys, mandatory from 1 September 2026. The price for the model is $12.50 per million input tokens and $75 per million output tokens, considerably higher than the standard GPT-5.6-Sol.
What does this mean for Dutch security professionals?
For most organisations, direct access is not relevant for the time being, Daybreak Red is designed for large security firms and specialised red teams, not for internal IT departments or individual penetration testers.
What is relevant is the broader shift this model represents. Offensive AI capabilities are being commercialised, admittedly behind an access gate, but the barrier is lowering. For defensive teams, this means that attack scenarios which were time-consuming and scarce last year are increasingly becoming automatically available to adversaries with fewer scruples about access rules.
Three practical implications:
Update threat modelling. Exploit-chain automation is no longer a theoretical risk. Teams maintaining threat models for sectors such as financial advice, healthcare or legal should explicitly include AI-driven attack paths.
Strengthen logging and detection. Systems that were well protected against manual attacks are not automatically resilient against faster, automated exploitation. Monitoring and anomaly detection become more urgent.
Check data processing agreements. If your organisation engages an external security firm that uses GPT-5.6-Cyber, you may be processing personal data through an OpenAI service. That requires a valid data processing agreement. OpenAI offers one for Enterprise and API customers, but the Daybreak-specific data processing agreement had not yet been publicly published at the time of launch. See our safety page for the current GDPR overview per model.
Astra as the next threshold
OpenAI is working on an internal model called Astra, aimed at long, autonomous tasks. SecurityWeek reports that internal evaluations suggest Astra may reach the "critical" risk threshold for autonomous cyberattacks, a classification OpenAI has not yet assigned to any published model. Whether and when Astra will be released, and under which access conditions, is currently unknown.
For a broader overview of models and their security profiles, including the GPT-5.6 family, see the models page.
Sources: SecurityWeek, 11 Aug 2026 · DataNorth AI · Forbes, 11 Aug 2026