OpenAI launches GPT-5.5-Cyber: a new round of the AI safety race amid Anthropic's restrictions
On June 22, OpenAI officially released the full version of its specialized model, GPT-5.5-Cyber, designed for professional vulnerability search, verification, and remediation. This move was a direct response to tightening regulatory restrictions on Anthropic, which created a unique market window for OpenAI.
The model is a key component of the Daybreak program and is not intended for mass users. Access to GPT-5.5-Cyber will only be granted to verified cybersecurity professionals who require advanced capabilities for conducting authorized defensive work.
Context: How Anthropic's Restrictions Opened the Door
On June 9, Anthropic launched two versions of its Claude models — Fable 5 and Mythos 5. However, by June 12, following a directive from the U.S. government under export controls, the company was forced to disable these models. This decision triggered a wave of problems: on June 23, law firm Legion filed a lawsuit against the U.S. government, claiming that losing access to Anthropic's models disrupted their document preparation tools.
Against this backdrop, OpenAI chose a more pragmatic path. The company pre-coordinated checks with U.S. federal agencies and opened access to GPT-5.5-Cyber exclusively to verified users, minimizing regulatory risks.
Technical Performance: What GPT-5.5-Cyber Shows
According to OpenAI, the new model delivers impressive results on specialized benchmarks. On the CyberGym test, GPT-5.5-Cyber scored 85.6% compared to 81.8% for standard GPT-5.5. On ExploitGym, it scored 39.5% versus 25.95%, and on SEC-bench Pro, it achieved 69.8% compared to 63.1%.
However, the picture is not so clear-cut. The UK AI Security Institute (AISI) conducted independent tests: GPT-5.5 completed the 32-step corporate attack simulation The Last Ones in 2 out of 10 attempts, while Mythos Preview succeeded in 3 out of 10. In a later update, AISI showed that the new version of Mythos Preview handled this scenario in 6 out of 10 attempts, for the first time also completing the second scenario, Cooling Tower.
Access Strategy and Partnerships
OpenAI has clearly stated that GPT-5.5-Cyber is a tool for a select few. For most clients, GPT-5.5 with Trusted Access for Cyber and Codex Security is still recommended. The company has also launched the Daybreak Cyber Partner Program, which includes giants such as Akamai, Check Point, Cisco, CrowdStrike, IBM, Palo Alto Networks, and others.
Notably, the Patch the Planet initiative for open-source projects, created in collaboration with Trail of Bits, deserves attention. Among the first participants are cURL, Go, Python, Sigstore, and pyca/cryptography. Since March, the cloud version of Codex Security has scanned over 30 million commits across more than 30,000 codebases, with human reviewers flagging over 70,000 findings as fixed.
My expert commentary: The AI security market is entering a new phase — moving from demonstrating capabilities to practical application. However, the gap between test results and real-world attack scenarios (as shown by AISI data) remains significant. The key question is not who scores higher on benchmarks, but whose models can work effectively in real corporate infrastructures with minimal false positives. OpenAI has bet on controlled access and partnerships, which appears to be a more mature approach than Anthropic's attempt to release "private" models without proper regulatory preparation.