GPT-5.6-Cyber: OpenAI Loosens Safeguards for Security Work

What happened

OpenAI has released GPT-5.6-Cyber, a version of its model tuned for offensive and defensive security work. Its defining feature is a deliberately reduced set of safeguards. According to reporting on the launch, the model completes 95% of advanced cyber requests and can identify zero-day vulnerabilities, meaning previously unknown flaws that vendors have not yet patched. 1 The reporting frames the release around exploit development, an area where general-purpose assistants have usually refused or hedged. 1

The release follows the GPT-5 generation, which OpenAI promoted mainly on benchmark gains. The company said GPT-5 set new state-of-the-art results across several domains. It reported 94.6% on AIME 2025 math without tools, 74.9% on SWE-bench Verified and 88% on Aider Polyglot for real-world coding, 84.2% on MMMU for multimodal understanding, and 46.2% on HealthBench Hard. 2 OpenAI also said GPT-5 pro, using extended reasoning, scored 88.4% on GPQA without tools. 2

Two very different launch messages

The two announcements are pitched in different ways. GPT-5 was presented as a broad capability upgrade, and OpenAI stressed that its benchmark gains carry over into everyday use. 2 GPT-5.6-Cyber is pitched less on what the model can do and more on what it is now allowed to do. The headline figure is a completion rate on sensitive requests, not a score on a public benchmark. 1

This distinction matters. A 95% completion rate on advanced cyber tasks describes how rarely the model refuses as much as how capable it is. 1 A model can be highly skilled and still decline most exploit-related prompts. A model with looser guardrails turns that latent skill into usable output. The strong coding results OpenAI reported for GPT-5, especially on SWE-bench Verified, suggest the underlying engineering ability was already substantial. 2 One plausible reading is that the cyber variant mainly unlocks that ability for a specific use, rather than adding a large new capability jump. That is an interpretation, not something OpenAI has confirmed.

Why it matters

Security has always been a hard case for AI safety policy. The skills that let a defender find and fix a vulnerability are the same ones an attacker uses to weaponize it. General-purpose models have usually settled this tension by refusing broadly, which frustrates legitimate penetration testers and vulnerability researchers. A purpose-built model with reduced safeguards is a different answer: split the problem off into its own product instead of tuning one model to serve everyone.

The claim that the model finds zero-day flaws is the most significant detail. 1 Zero-day discovery is valuable to defenders because it lets them patch before attackers act. The same output in the wrong hands is an attack tool. How much benefit this brings, and how much risk, likely depends on factors such as who gets access, how use is monitored, and what happens when the model finds a vulnerability. The available reporting does not settle those questions. Readers should be wary of assuming any particular access-control regime until OpenAI documents one.

Context and caveats

The two sets of figures are not directly comparable. GPT-5's numbers come from named, public benchmarks that third parties can in principle reproduce. 2 The 95% completion figure for GPT-5.6-Cyber is described in terms of "advanced cyber requests." 1 What counts as an advanced request, and how completion was judged, are not detailed. Until there is an evaluation methodology, the number shows how permissive the model is more clearly than how well it performs.

The naming also shows a shift in how OpenAI is packaging its models. GPT-5 and GPT-5 pro were presented as tiers of one general system, separated mainly by reasoning depth. 2 GPT-5.6-Cyber points toward domain-specific variants whose policies, not just capabilities, differ from the flagship.

The takeaway

GPT-5.6-Cyber looks less like a raw capability milestone and more like a policy experiment. It tests whether a frontier lab can safely give a less-restricted model to the security field. The GPT-5 benchmarks suggest the technical foundation is strong. 2 The cyber variant's reported completion rate and zero-day findings show OpenAI is now willing to point that foundation directly at exploitation tasks. 1 Whether this proves to be a net gain for defenders will depend far more on deployment controls than on any single headline percentage. Those controls deserve the closest scrutiny as more details come out.