OpenAI launches GPT-5.5-Cyber: A new era of cybersecurity amid regulatory pressure on Anthropic
On June 22, OpenAI officially unveiled the full version of its latest model — GPT-5.5-Cyber. This is not just another update to a language model, but a specialized tool designed to find, verify, and fix software vulnerabilities. The release came at an extremely tense time for the industry — right after the U.S. government imposed export restrictions that forced Anthropic to withdraw its advanced models, Fable 5 and Mythos 5.
A Window of Opportunity for OpenAI
The situation surrounding Anthropic has created a unique "window" for OpenAI. On June 9, Anthropic released Claude Mythos 5 — a model with relaxed restrictions intended for professional use. However, on June 12, under pressure from a U.S. government directive related to export controls, the company was forced to disable access to these models. This sparked a wave of discontent, including a lawsuit from the law firm Legion, whose development team is based in Canada.
OpenAI, by contrast, took a path of proactive cooperation with regulators. The company emphasizes that all model checks were pre-coordinated with U.S. federal agencies, and access to GPT-5.5-Cyber is granted exclusively to verified cybersecurity professionals. This allowed OpenAI to avoid regulatory risks and fill the vacated niche.
Technical Specifications and Benchmarks
GPT-5.5-Cyber demonstrates impressive results on specialized tests. On the CyberGym benchmark, the model achieved a score of 85.6%, significantly higher than the 81.8% of the standard GPT-5.5. Even more telling is the ExploitGym test, which assesses the ability to turn a vulnerability into a working exploit: here, the new model's result is 39.5% compared to 25.95% for the regular version. On the long-term vulnerability search test SEC-bench Pro, the model scored 69.8% versus 63.1%.
However, as data from the UK AI Security Institute (AISI) shows, the picture is mixed. In a simulation of a complex corporate attack called "The Last Ones," GPT-5.5-Cyber succeeded in 2 out of 10 attempts, while the withdrawn Mythos Preview from Anthropic succeeded in 3 out of 10. Moreover, after an update, Mythos Preview achieved a result of 6 out of 10, and GPT-5.5 achieved 3 out of 10. This suggests that, although OpenAI's specialized model outperforms the base version, the arms race in AI security is far from over.
Ecosystem and Access Restrictions
GPT-5.5-Cyber is just one part of the broader Daybreak initiative. OpenAI has also launched a partner program involving giants such as Akamai, Cisco, CrowdStrike, and IBM, as well as the Patch the Planet initiative for open-source projects (cURL, Go, Python). According to the company, since March, the Codex Security cloud service has scanned over 30 million commits, identifying more than 70,000 fixed vulnerabilities.
The model is not intended for mass use. OpenAI has made it clear that this is a tool for select professionals working in authorized scenarios. For regular customers, the standard GPT-5.5 with Trusted Access for Cyber is still recommended.
Expert Opinion. The launch of GPT-5.5-Cyber amid Anthropic's restrictions is a masterful move by OpenAI, demonstrating an understanding of regulatory dynamics. However, AISI data shows that in the area of autonomous attacks, Anthropic's model may still be ahead. The key question now is not who can find a vulnerability faster, but who can more effectively and safely integrate these capabilities into real corporate systems without creating new risks. The cybersecurity market is entering an era where AI agents become both the primary weapon and the primary shield.