GPT-5.5-Cyber exits preview with 85.6% CyberGym score, outpaces Mythos
OpenAI moved GPT-5.5-Cyber out of preview this week, opening access to security partners and individual researchers. The model edges out Mythos on the CyberGym vulnerability-detection benchmark.
OpenAI's GPT-5.5-Cyber exited preview on June 23 and is now available to cybersecurity partners and individual researchers. The model scored 85.6 percent on CyberGym, the primary benchmark for vulnerability detection in code, topping Mythos's 83.8 percent. OpenAI announced the move in a blog post titled "Daybreak: Securing the World," positioning the release as part of a broader push into enterprise security tooling.
The partner list includes non-US firms, a detail that may draw regulatory scrutiny given existing export controls on advanced AI capabilities. Anthropic faced similar questions when Claude models were made available to international security vendors. The inclusion of foreign entities raises questions about how OpenAI will navigate dual-use technology restrictions, particularly as the US government has tightened controls on AI systems capable of discovering zero-day exploits. CyberGym scores measure a model's ability to flag common vulnerability patterns across multiple programming languages; the 1.8-point gap between GPT-5.5-Cyber and Mythos represents roughly 12 additional correct identifications in the 680-sample test set. Both models are closed-weight and API-only, distinguishing them from open-weight code models like DeepSeek-Coder and CodeLlama, which security researchers can fine-tune locally but which lack the specialized vulnerability-detection training that GPT-5.5-Cyber received.




