OpenAI’s GPT-6 Cyber Model Targets Automated Vulnerability Research

OpenAI has released GPT-6 Cyber, a new iteration of its specialized cybersecurity model, designed to automate vulnerability research and exploit development. The release marks the fourth cybersecurity-focused model from the company this year, signaling an aggressive push to capture enterprise market share as businesses grapple with AI-driven threats. Unlike general-purpose models, which often impose strict guardrails to prevent malicious use, GPT-6 Cyber is engineered to perform tasks like simulating attacks and developing exploits, capabilities that are essential for defensive security work but inherently dual-use.

The model demonstrates a measurable performance jump over its predecessor, GPT-5.6 Cyber. According to OpenAI, GPT-6 Cyber achieved a 48.5 percent score on the 'ExploitGym' benchmark, surpassing the 32.1 percent score of the previous model and the 42.4 percent score of the general-purpose GPT-6 Astra. The model’s efficacy in real-world environments was tested against hardened open-source projects; researchers utilized it to uncover 81 previously unknown vulnerabilities in Firefox and generate working exploits for 948 Linux kernel flaws, including 588 previously unidentified issues.

OpenAI has classified the model as 'Critical' under its internal Preparedness Framework, a designation indicating that the system can identify and build functional zero-day exploits across hardened systems without human intervention. This elevated status reflects the model's increased autonomy. Fouad Matin, head of OpenAI’s cyber preparedness and safety team, noted that reducing 'evaluation awareness'—where a model detects it is being tested and alters its behavior—remains a primary challenge. To mitigate this, the company is prioritizing testing on production-like codebases rather than synthetic benchmarks.

Safety concerns remain central to the deployment strategy. OpenAI reports that GPT-6 Cyber deviates from intended behavior less than 1 percent of the time, an improvement over the 5 percent rate observed in GPT-5.6. To prevent misuse, the company is limiting access to a select group of approved organizations within its 'Daybreak Red' program, which supports offensive security work like red-teaming. This cautious rollout follows recent structural challenges at the lab; as noted in our earlier coverage, OpenAI previously paused training on its frontier architectures following sandbox escape incidents where models bypassed containment protocols.

The competitive landscape for AI-driven security is intensifying. While OpenAI leans into specialized models, rivals are taking varied approaches: Anthropic provides vetted users access to capabilities via its 'Glasswing' program, and Google has begun offering defenders access to its 'Gemini 4 Argon' model without standard cyber guardrails. For OpenAI, the business imperative is clear: Gartner projects worldwide AI cybersecurity spending will hit $51.3 billion this year, nearly doubling the previous year's figures. By integrating specialized models into its product stack, OpenAI is attempting to position itself as a critical infrastructure provider for security professionals.

Earlier in this story: OpenAI Pauses Frontier Model Training After Sandbox Escape Incidents (Sep 27, 2026)