OpenAI introduces GPT-5.5-Cyber: a response to Anthropic's limitations and a new turn in the AI safety race.
On June 22, OpenAI officially launched the full version of its specialized model, GPT-5.5-Cyber, designed for finding, verifying, and fixing software vulnerabilities. This release came amid severe restrictions imposed on Anthropic's models, creating a unique market window for OpenAI.
The model is a key component of the Daybreak initiative, OpenAI's cyber defense program. It is important to emphasize: GPT-5.5-Cyber is not intended for the mass user. Access will only be granted to verified cybersecurity professionals who need enhanced tools for authorized defensive work.
How the "Window of Opportunity" Emerged
On June 9, Anthropic opened access to two versions of its Claude models: Fable 5 and the "private" Claude Mythos 5 with relaxed restrictions. However, on June 12, after receiving a directive from the U.S. government under export controls, Anthropic was forced to disable these models. This decision caused a wave of problems: on June 23, the company Legion filed a lawsuit against the U.S. government, claiming that losing access to Anthropic's models disrupted its tools for preparing legal documents.
OpenAI chose a fundamentally different path. The company coordinated its checks with U.S. federal agencies in advance, then opened GPT-5.5-Cyber exclusively to verified users. This is a strategic move demonstrating how to balance innovation with regulatory requirements.
What GPT-5.5-Cyber Can Do: Numbers and Facts
Test results are impressive. On the CyberGym benchmark, the model scored 85.6% compared to 81.8% for standard GPT-5.5. An even larger gap was seen on ExploitGym: 39.5% versus 25.95%. On SEC-bench Pro, results were 69.8% versus 63.1%. These tests evaluate the model's ability to turn vulnerabilities into working exploits and conduct long-term proof-of-concept searches.
Of particular interest is the comparison with Anthropic Mythos 5. On CyberGym, GPT-5.5-Cyber outperformed the competitor: 85.6% versus 83.8%. However, data from the UK AI Security Institute (AISI) paints a more complex picture. In the corporate attack simulation The Last Ones, GPT-5.5-Cyber completed the scenario in 2 out of 10 attempts, while Mythos Preview did so in 3 out of 10. Later, AISI updated the data: a new version of Mythos Preview passed the test in 6 out of 10 attempts, and GPT-5.5 in 3 out of 10.
Access Strategy and Partnerships
OpenAI has clearly restricted access to the model. For most clients, GPT-5.5 with Trusted Access for Cyber and Codex Security is still recommended. In parallel, the Daybreak Cyber Partner Program has been launched. Participants include Akamai, Check Point, Cisco, CrowdStrike, IBM, Palo Alto Networks, Proofpoint, SentinelOne, Wiz, Zscaler, and other security giants.
Another important direction is the Patch the Planet initiative for open-source projects, created jointly with Trail of Bits and with participation from HackerOne. Early participants include cURL, Go, Python, Sigstore, and pyca/cryptography. According to the company, since March, the cloud version of Codex Security has scanned over 30 million commits in more than 30,000 code bases, with human reviewers noting over 70,000 findings as fixed.
My comment: The release of GPT-5.5-Cyber is not just a technological step, but a clear signal to the market. OpenAI demonstrates a willingness to work within regulatory constraints while gaining a competitive advantage. However, AISI data shows that the AI security race is far from over: different models show varying effectiveness in different scenarios. The key question remains not just model performance, but its ability to adapt to real-world, rather than laboratory, attack conditions. The zero-day vulnerability market is in for a major transformation.