OpenAI Releases Specialized GPT-5.6-Cyber to Automate Threat Recognition and Analysis

OpenAI has introduced a new version of its flagship AI, GPT-5.6-Cyber. The model is designed specifically for cybersecurity professionals and trained to discover vulnerabilities, build exploit chains, and test systems for resilience. 

The company emphasizes that launching this tool is a response to cybercriminals increasingly leveraging AI to orchestrate autonomous attacks.

New Access Level and Reduced Failures

The primary distinction between the new model and the standard GPT-5.6 Sol is a intentionally lowered failure threshold for "grey area" prompts, such as developing defense bypass methods or privilege escalation. 

According to OpenAI's internal metrics, the specialized version successfully fulfills up to 95% of such requests, whereas the base model responds to only 1,5–2% of identical prompts due to strict system guardrails.

First Zero-Day Vulnerabilities Discovered

The efficiency of these algorithms has already been demonstrated in practice. During initial testing, GPT-5.6-Cyber uncovered a series of critical flaws in widely used software:

• 2 zero-day vulnerabilities in Google Chrome's V8 JavaScript engine.
• 5 vulnerabilities in a mobile OS, including a complete privilege escalation chain.
• Over 400 security issues in the kernel of a popular OS and three critical flaws in a widely used database.

OpenAI notes that this AI can perform complex exploit constraint modeling within 24 hours, which previously took humans weeks to complete.

Strict Control and Security Measures

Under OpenAI's Preparedness Framework, the model reached a "High" cybersecurity capability level, requiring elevated risk-management measures. OpenAI stresses that access is restricted strictly to verified organizations with legal identity verification and the mandatory adoption of hardware security keys. 

To mitigate risks, developers recommend running AI agents exclusively in isolated environments without access to the internet or production networks.