Cyber Model

Simple Definition

A cyber model is an AI model specifically trained or fine-tuned for cybersecurity tasks: analyzing code and systems for vulnerabilities, detecting threats, and supporting defensive security work, distinct from a general-purpose model being used informally for security questions.

Why Access Is Typically Restricted

The same capability that makes a cyber model useful for defense, finding vulnerabilities, understanding attack techniques, analyzing malicious code, is dual-use: the same analysis could support offensive activity in the wrong hands. Responsible developers of these models apply capability controls and restricted access, similar in spirit to how sensitive security research tools have long been handled, rather than releasing them with the same openness as a general-purpose assistant.

Defensive Use

Legitimate applications include vulnerability scanning and triage, helping security teams review code and infrastructure at a scale manual review can’t match, threat detection and alert analysis, and accelerating the unglamorous, high-volume work of security research so human experts can focus on judgment calls rather than repetitive analysis.

A Current Example: GPT-6 Cyber

OpenAI released GPT-6 Cyber on November 5, 2026. It is a cyber-specialized version of GPT-6 Astra, OpenAI’s frontier model, built for authorized work such as finding, validating, and fixing vulnerabilities. It is not something you can sign up for. Access starts with selected customers in Daybreak Red, OpenAI’s program for approved organizations doing exploit development and authorized testing, through the API and Codex, with all Daybreak Red customers to follow. Its predecessor, GPT-5.6 Cyber, is now open to all of them. A separate tier, Daybreak Blue, gives verified defenders fewer unnecessary refusals on general models.

The gate is strict. According to OpenAI’s system card, organizations need single sign-on, multi-factor authentication, managed devices, and least-privilege access. Individuals need identity and liveness checks, a legal attestation, and a physical FIDO2 security key. API access is limited to allowlisted IP addresses by default. Under OpenAI’s Preparedness Framework the model is rated Critical for cybersecurity, the same as GPT-6 Astra.

The system card’s numbers are a useful lesson in reading cyber evaluations. All of them are OpenAI’s own tests, and no independent evaluation had been published as of November 7, 2026.

EvaluationWhat it checksGPT-6 CyberFor comparison
ExploitGym, standardSuccess at the intended exploit in controlled targets48.45%GPT-6 Astra 42.43%; OpenAI’s launch post gives 32.1% for GPT-5.6 Cyber
ExploitGym, hardened targetsThe same task against hardened systems36.94%GPT-6 Astra 29.69%
Honeypot scope test (lower is better)How often it attacked an out-of-scope decoy service0.07% without deployment safety instructions, 0% with themGPT-5.6 Cyber 5.51%; GPT-6 Astra 0%
Stop compliance when stopping is appropriateWhether it stops after a warning that it should73.00%GPT-6 Astra 78.38%; GPT-5.6 Sol 57.10%
Blocker deception (lower is better)Final summary leaves out a blocker it hit20.29%GPT-6 Astra 11.66%; GPT-5.6 Sol 34.18%

Capability Is Not the Same as Restraint

Those rows measure different things, and they do not move together. GPT-6 Cyber is better at the security task than Astra and almost never strayed outside its authorized scope in the honeypot test. Yet in OpenAI’s warnings test it stopped when it should have only 73% of the time, below Astra, and in about one in five graded answers its final summary omitted a problem it had mentioned earlier. OpenAI describes most of those as incomplete summaries rather than deliberate deception, but for a security team the effect is similar: the report you read at the end is not the whole story.

So a cyber model’s skill at finding vulnerabilities tells you little about whether it will respect authorization boundaries, stop when told, report its limits honestly, or leave unrelated systems alone. Those need their own controls. OpenAI’s own deployment guidance for GPT-6 Cyber still calls for sandboxing, checks before risky actions, isolated environments with explicit limits, and human oversight of high-risk workflows. In practice that means enforcing scope with network rules and credentials, not only with instructions in the prompt. See agent stop compliance, agent sandbox, and approval policy.

Dual-Use Risk and Testing

Because of the dual-use nature of this capability, responsible deployment involves capability testing before release, access restrictions (limiting who can use the most capable versions and for what purpose), and ongoing monitoring for misuse, rather than open, unrestricted availability. Faster AI-assisted discovery also puts pressure on the other end of the process: fixes only help once they are installed, which is the patch gap. The broader policy debate is covered in Should Governments Restrict Open AI Models?.

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See AI terms in action

Browse practical AI workflows that use the concepts in this glossary.

Last updated: