On Thursday, OpenAI announced the limited release of the GPT-5.5-Cyber preview version to verified security teams, a specialized version of its latest model, GPT-5.5, tailored for cybersecurity. The company emphasized that this version is not intended to enhance offensive or defensive cyber capabilities, but rather to relax the built-in restrictions of the model when handling security tasks through targeted training, allowing authorized teams to more efficiently perform workflows such as vulnerability identification, patch validation, and malware analysis. The general version, due to its security protection mechanisms, makes such tasks more difficult.
This move follows in the footsteps of competitor Anthropic, which released the Claude Mythos preview version a month ago. As part of the "Glasswing" (Project Glasswing) cybersecurity initiative, the latter is also only available to specific enterprises. The simultaneous launch of these specialized models reflects how major model developers are shifting from competition over general capabilities to precise adaptation in vertical scenarios.
Notably, these developments have drawn intense attention from U.S. government officials. Fed Chair Powell and Treasury Secretary Beyenert recently discussed the potential impact of the Mythos model with top bank CEOs; Vice President Vance also held phone calls with tech giants. Despite being placed on a blacklist by the Pentagon, Anthropic's CEO continued to communicate with high-level officials in the Trump administration regarding the model's capabilities.
As large models penetrate high-sensitivity fields such as finance and security, the "relaxed restrictions" strategy of specialized versions improves professional efficiency but also brings new governance challenges. Industry observers believe that establishing a dynamic balance between capability release and risk control will become a key issue in the next phase of large model deployment and the coordinated evolution of safety compliance.
