OpenAI has launched GPT-5.5-Cyber, a specialized AI model for cybersecurity professionals, amid Anthropic's restrictions.
OpenAI has officially launched GPT-5.5-Cyber, a specialized version of its AI model aimed at enhancing cybersecurity measures. This rollout comes amid ongoing restrictions affecting Anthropic's AI models.
Unlike typical consumer models, GPT-5.5-Cyber is exclusively available to certified cybersecurity experts, ensuring that its advanced capabilities are utilized in authorized contexts. This model is part of OpenAI's Daybreak initiative, which focuses on strengthening cybersecurity tools.
Context of the Launch
The opportunity for this launch arose after Anthropic unveiled two versions of its Claude model, which were later restricted by the U.S. government due to export controls. This led to legal challenges, including a lawsuit filed by Legion, a San Jose-based organization, against the U.S. government for limiting access to Anthropic's models, which affected its operations in legal document preparation.
In contrast, OpenAI ensured compliance with federal regulations prior to the launch of GPT-5.5-Cyber, thereby avoiding similar complications.
Performance Metrics
According to OpenAI, GPT-5.5-Cyber achieved an impressive score of 85.6% on the CyberGym benchmark, outperforming the standard GPT-5.5, which scored 81.8%. This benchmark evaluates a model's ability to replicate known vulnerabilities within controlled software environments.
Further tests showed that GPT-5.5-Cyber scored 39.5% on ExploitGym and 69.8% on SEC-bench Pro, indicating its strong capabilities in converting vulnerabilities into exploits and discovering long-term vulnerabilities, respectively.
OpenAI's Daybreak program aims not only to identify vulnerabilities but also to expedite their resolution. Since March, the cloud version of Codex Security has scanned over 30 million code commits, resulting in more than 70,000 vulnerabilities being marked as resolved.
Access Limitations and Partnerships
Access to GPT-5.5-Cyber is restricted to vetted users, with OpenAI recommending its standard GPT-5.5 model with Trusted Access for most clients. The company has also initiated the Daybreak Cyber Partner Program, collaborating with notable cybersecurity firms such as Cisco, IBM, and CrowdStrike.
Additionally, OpenAI has launched the Patch the Planet initiative in collaboration with Trail of Bits, focusing on enhancing cybersecurity for open-source projects. This program includes participation from well-known projects like Python and cURL.
Experts have warned that advanced AI models have the potential to significantly alter the landscape of cybersecurity, enabling quicker attacks and reducing the time between discovering and exploiting vulnerabilities.
In a related development, Anthropic recently discussed the safe deployment of AI agents in corporate settings, highlighting the risks and necessary precautions for effective cybersecurity.
