OpenAI split its Daybreak program for security professionals into two tiers and released GPT-5.6-Cyber, a model trained for vulnerability research and exploit work. The blue tier lifts standard safety filters on general models, while the red tier gives the specialised model to approved researchers doing authorised testing. Access requires identity checks, approved-use limits and, from September, hardware security keys.
What changed
Security researchers using OpenAI's general models hit safety blocks on legitimate offensive testing work, and the earlier cyber model still refused many requests.
What it unlocks
Approved security teams can ask a model directly for exploit code, authentication bypasses and privilege-escalation work that general models decline.
- 95.0% request completion rate
- 1.5% for GPT-5.6 Sol
- 57.3% for GPT-5.5-Cyber
- 400+ kernel privilege-escalation flaws
What you need to act on it
- approval into the Daybreak program
- identity verification and legal attestations
- hardware security keys for individual accounts from 1 September 2026
Sources