Archive · 11 Aug 2026
OpenAI unveils GPT-5.6-Cyber with fewer guardrails for trusted security researchers
OpenAI is a company that makes AI, which are smart computer programs.
Some of these programs are very good at cybersecurity, which means keeping computers safe from hackers.
OpenAI made a special new version of its model, called GPT-5.6-Cyber, to help the 'good guys' who protect computer systems.
The regular version of the model only answered about 1.5 out of every 100 hard security questions, but the new version answered 95 out of 100.
It is only given to trusted people and companies doing approved security work.
OpenAI is also expanding its Daybreak program into two groups: Daybreak Blue for most defenders and Daybreak Red for experts who study hacking methods.
Big companies like IBM and Cisco can now put these tools into their own security products.
This comes just days after OpenAI delayed a different model called Astra because it became too good at hacking during safety tests.
OpenAI says it still has extra checks, like identity verification and monitoring, to stop bad people from using these tools.
OpenAI introduced GPT-5.6-Cyber, a more permissive cyber model for vetted security defenders.
Its Daybreak programme now has two tiers: Daybreak Blue for GPT-5.6 Sol without system-level guardrails and Daybreak Red for GPT-5.6-Cyber.
In internal tests, GPT-5.6-Cyber completed 95% of advanced cybersecurity requests versus 1.5% for GPT-5.6 Sol and 2% for Daybreak Blue access.
The launch follows the delay of OpenAI's Astra model after safety testing found it reached critical hacking capabilities.
Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks can incorporate the models into security products and managed services.
- Who
- OpenAI, and the vetted cybersecurity defenders it serves, including partners Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks.
- What
- OpenAI unveiled GPT-5.6-Cyber and expanded its Daybreak programme, giving trusted defenders access to cyber models with fewer guardrails.
- When
- No specific date was given; the announcement came days after OpenAI delayed the release of its Astra model.
- Where
- Not specified in the articles.
- Why
- To equip trusted defenders with more powerful cybersecurity tools ahead of possible autonomous cyberattacks while keeping capabilities out of malicious hands.
- New model introduced
- GPT-5.6-Cyber, a cyber-permissive model for trusted defenders
- Related development
- OpenAI delayed its Astra model after safety testing found critical hacking capabilities
- Daybreak tiers
- Daybreak Blue (GPT-5.6 Sol with reduced guardrails) and Daybreak Red (GPT-5.6-Cyber)
- GPT-5.6-Cyber completion rate
- 95% of advanced cybersecurity requests in internal tests
- GPT-5.6 Sol completion rate
- 1.5% of such requests
- GPT-5.5-Cyber completion rate
- 57.3% of such requests
- Partner companies
- Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks
- Access controls
- Identity verification, account security, monitoring, use restrictions and legal attestations