OpenAI introduces GPT‑5.5‑Cyber for high-impact cybersecurity research

OpenAI Group PBC has developed a version of GPT-5.5 that is specifically optimized for cybersecurity research.

GPT‑5.5‑Cyber, as the model is called, made its debut on Thursday. It’s available in limited preview through a program called Trusted Access for Cyber, or TAC, that OpenAI launched in February. The initiative gives cybersecurity researchers expanded access to the company’s algorithms.

GPT-5.5, OpenAI’s newest and most capable large language model, launched last month. The algorithm is a highly capable programming assistant: it helped OpenAI develop software that significantly sped up some of its server clusters. The mechanisms that make GPT-5.5 adapt at code generation can also be harnessed to find cybersecurity vulnerabilities.

OpenAI limits access to the model’s cybersecurity capabilities for most users to prevent misuse. TAC, the threat research initiative that it launched in February, eases the usage restrictions for cybersecurity experts. The GPT‑5.5‑Cyber model that debuted on Thursday supports an even broader range of cybersecurity use cases than the standard TAC version of GPT-5.5.

If a user with a regular ChatGPT account asks GPT-5.5 to exploit a vulnerable website, the model either refuses the request or assumes that the user needs help applying a fix. In the latter scenario, the model generates a series of vulnerability remediation suggestions. 

Researchers who participate in OpenAI’s TAC program receive more detailed prompt responses. GPT-5.5 can generate a technical description of how hackers might go about launching a cyberattack against a vulnerable system. The model can even generate sample malware, but it doesn’t validate that the exploit works.

The new GPT-5.5-Cyber model goes a step further. It can not only generate a vulnerability exploitation plan but also validate that it works by launching a simulated cyberattack against the system being studied. That makes it useful for tasks such as automating red teaming exercises. Those are projects in which a company tests the security of its infrastructure by mimicking hacker activity.

OpenAI says that GPT-5.5-Cyber achieved a 81.9% score on CyberGym, a benchmark that measures LLMs’ cybersecurity capabilities. It comprises more than 1,500 historical vulnerabilities from hundreds of open-source projects. 

OpenAI has integrated GPT-5.5-Cyber’s expanded threat research capabilities with new safety guardrails. The company will use “stronger verification” to prevent hackers from gaining access to the model. In addition, OpenAI has implemented an enhanced set of misuse monitoring features to ensure that researchers who are given access adhere to cybersecurity best practices. 

Initially, OpenAI will only make GPT-5.5-Cyber available to a limited number of “defenders responsible for securing critical infrastructure.” The company says that the standard version of GPT-5.5 that is available through its TAC program is a better option for most cybersecurity projects. 

GPT-5.5-Cyber may create more competition for Claude Mythos Preview, Anthropic PBC’s most advanced LLM. The latter algorithm is also adept at finding software vulnerabilities. Earlier this year, Anthropic made it accessible to a limited number of organizations to help them improve their cybersecurity posture.

Image: OpenAI