OpenAI launches GPT-5.5-Cyber amid tightening control over Anthropic: A new era of AI cybersecurity
On June 22, OpenAI officially unveiled the full version of its specialized model, GPT-5.5-Cyber, designed for finding, verifying, and fixing vulnerabilities. This release is a direct response to recent restrictions imposed on Anthropic's models and marks a new stage in the race for dominance in AI security.
The model, a key element of the Daybreak program, will not be available to the general public. Its target audience is verified cybersecurity experts who require advanced tools for conducting authorized defensive operations. OpenAI has chosen a path of proactive alignment with US federal agencies, unlike Anthropic, whose Fable 5 and Mythos 5 models were disabled after receiving a government directive under export controls.
Technical Metrics and Comparative Analysis
Benchmark results demonstrate significant progress. On the CyberGym test, which includes 1,507 tasks based on real vulnerabilities in 188 open-source projects, GPT-5.5-Cyber achieved a score of 85.6%, surpassing the standard GPT-5.5 (81.8%) and even the blocked Anthropic Mythos 5 model (83.8%).
Even more impressive results were achieved on other tests. On ExploitGym, which evaluates the ability to turn a vulnerability into a working exploit, the model scored 39.5% compared to 25.95% for the regular version. On SEC-bench Pro, which simulates long-term vulnerability hunting, the result was 69.8% versus 63.1%.
However, as data from the UK AI Security Institute (AISI) shows, the picture is less clear in more complex full attack simulation scenarios. In The Last Ones test, simulating a multi-stage corporate attack, GPT-5.5-Cyber succeeded in 3 out of 10 attempts, while the updated Mythos Preview version managed 6 out of 10 attempts. This highlights that, despite its narrow specialization, the general agentic capabilities of Anthropic's models remain higher in some aspects.
Access Strategy and Partnerships
OpenAI has clearly defined the usage boundaries for GPT-5.5-Cyber. The model is intended exclusively for verified users in authorized scenarios. For other clients, the company recommends using GPT-5.5 with Trusted Access for Cyber and Codex Security features.
Under the Daybreak Cyber Partner Program, OpenAI has already attracted giants such as Akamai, Check Point, Cisco, CrowdStrike, IBM, Palo Alto Networks, Proofpoint, SentinelOne, Wiz, and Zscaler. Additionally, the Patch the Planet initiative for open-source projects has been launched, created jointly with Trail of Bits, HackerOne, and other partners. Among the first participants are cURL, Go, Python, Sigstore, and pyca/cryptography.
My analysis: The launch of GPT-5.5-Cyber is not just a technological breakthrough but a strategic move amid tightening regulatory pressure. OpenAI demonstrates that it can offer powerful cybersecurity tools without violating export restrictions, unlike Anthropic. However, the performance gap in complex agentic scenarios indicates that the race to create the "perfect" AI hacker is far from over. In the coming months, the key factor will be not only the raw power of the models but also the ability of companies to build safe and legitimate channels for their use.