OpenAI launches GPT-5.5-Cyber: a new era of cybersecurity amid Anthropic restrictions
On June 22, OpenAI officially unveiled the full version of GPT-5.5-Cyber, a specialized model designed for finding, verifying, and fixing vulnerabilities. This move came amid severe restrictions imposed on Anthropic, creating a unique market window for OpenAI.
Context: How OpenAI Seized the Moment
On June 9, Anthropic launched two versions of the Claude family models: Fable 5 and Mythos 5. The latter was positioned as a "private" base model with relaxed restrictions. However, on June 12, following a directive from the U.S. government under export control regulations, Anthropic was forced to disable these models. This sparked a wave of discontent: 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.
Unlike Anthropic, OpenAI chose a more cautious path. The company pre-coordinated all checks with U.S. federal agencies and opened GPT-5.5-Cyber exclusively to verified cybersecurity professionals. This approach allowed it to avoid regulatory risks and fill the vacated niche.
Technical Specifications: What the New Model Can Do
According to OpenAI, GPT-5.5-Cyber delivers impressive results on specialized benchmarks. On the CyberGym test, the model scored 85.6% compared to 81.8% for standard GPT-5.5. This benchmark, developed by UC Berkeley researchers, includes 1,507 tasks based on real vulnerabilities in 188 open-source projects. On ExploitGym, the model achieved 39.5% versus 25.95% for the regular version, and on SEC-bench Pro, it scored 69.8% compared to 63.1%.
Particularly noteworthy is the comparison with a competitor. GPT-5.5-Cyber outperformed Anthropic Mythos 5 on CyberGym: 85.6% versus 83.8%. However, in more complex scenarios, the situation is mixed. The UK AI Security Institute (AISI) reported that in a 32-step simulation of a corporate attack called The Last Ones, GPT-5.5-Cyber successfully completed the scenario in 2 out of 10 attempts, while Mythos Preview did so in 3 out of 10. Later, AISI updated the data: the new version of Mythos Preview passed the test in 6 out of 10 attempts, while GPT-5.5-Cyber only managed 3 out of 10.
Access Strategy and Partnerships
OpenAI has clearly stated that GPT-5.5-Cyber is not intended for mass users. The model is only available to verified professionals who require advanced capabilities for authorized defensive work. For regular customers, the company recommends standard GPT-5.5 with Trusted Access for Cyber and Codex Security.
Under the Daybreak program, OpenAI launched the Cyber Partner Program, which includes 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 was launched in collaboration with Trail of Bits and with participation from HackerOne. Early participants include cURL, Go, Python, Sigstore, and pyca/cryptography.
My expert opinion: The launch of GPT-5.5-Cyber is not just a technological breakthrough but a strategic move amid tightening AI regulation. OpenAI demonstrates how to effectively balance innovation and compliance, while competitors face regulatory constraints. However, the performance gap in complex attack scenarios (The Last Ones) shows that the cyber capabilities race is far from over. In the coming months, we will see how this dynamic impacts the entire cybersecurity market.