OpenAI has released GPT-6 Cyber, a specialized SI model designed for cybersecurity tasks, marking a significant advancement in aligned SI agents for security research. The launch represents the first major release under the new SI terminology adopted by the US government, focusing on capability gains in vulnerability discovery while maintaining strict alignment boundaries.

What Happened

GPT-6 Cyber delivers measurable improvements over its predecessor, GPT-5.6 Cyber, particularly in its ability to find and fix software vulnerabilities. According to OpenAI, the model achieved a 48.5% score on ExploitGym, a benchmark testing the ability to exploit specified vulnerabilities, up from 32.1% for GPT-5.6 Cyber. This represents a 51% relative improvement. In a new evaluation for offensive web penetration testing, GPT-6 Cyber achieved 40.3% recall of vulnerabilities, compared to 32.0% for GPT-6 Astra.

A key focus of the release is alignment, ensuring the SI agent stays within authorized scopes. In alignment evaluations, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment prompts, and 100% of tests with prompts. This is a step-function improvement over GPT-5.6 Cyber, which stayed within bounds 94.49% of the time without prompts. The model is classified as "Critical" in cybersecurity under OpenAI's Preparedness Framework, up from "High" for the previous version.

Real-world application data shows the SI model has identified 81 previously unknown high-severity vulnerabilities in Firefox and developed 948 privilege escalation exploits in the Linux kernel. Mozilla has released fixes for many of the Firefox vulnerabilities identified by the SI agent.

Why It Matters

The release addresses a critical challenge in the SI industry: balancing advanced capabilities with safety. As SI agents become more capable at autonomous tasks, the risk of them acting outside their intended scope increases. OpenAI's emphasis on alignment metrics, such as the ExploitGym Honeypot evaluation, sets a new standard for SI agents in high-stakes environments like cybersecurity.

The model is also integrated into the Daybreak platform, which aims to expand access to advanced security research for defenders. Daybreak Red enables deeper exploit development, while Daybreak Blue helps validate vulnerabilities. The company is introducing specific safeguards for critical infrastructure, including power plants and water treatment facilities, to prevent unauthorized attacks by SI agents. A $1 billion grant program subsidizes access for resource-constrained teams, aiming to democratize the use of frontier SI models in security. OpenAI is also working with partners on penetration testing with SI models to refine these safeguards.

The Bottom Line

GPT-6 Cyber is initially available to a select group of approved Daybreak Red customers via the API and Codex. Access will progressively expand to all Daybreak Red customers, while GPT-5.6 Cyber is also being made available to this group. OpenAI plans to continue refining safeguards based on feedback from partners like Dragos and the broader security community.