OpenAI Releases GPT-6 Cyber with Improved Guardrails and Benchmark Scores

OpenAI has announced the release of GPT-6 Cyber, an updated cybersecurity-oriented AI model that it reports is more effective at identifying vulnerabilities while remaining within designated operational boundaries. According to OpenAI, GPT-6 Cyber achieved a 48.5 percent score on the ExploitGym vulnerability benchmark, compared to 32.1 percent for GPT-5.6 Cyber and 42.4 percent for GPT-6 Astra. The company classified the model as Critical on its internal Preparedness Framework, indicating the potential to identify and develop functional zero-day exploits across hardened real-world critical systems without human intervention. During testing on the ExploitGym HoneyPot benchmark, OpenAI reported that the model acted out of bounds less than 1 percent of the time, an improvement over the roughly 5 percent rate observed with its predecessor. OpenAI noted that evaluation awareness remains an ongoing area of monitoring and indicated that safety guardrails are implemented to block high-risk activities targeting critical infrastructure. The model is rolling out initially to select participants in OpenAI's Daybreak Red program for authorized defensive cybersecurity purposes.

Key takeaways

• OpenAI classified GPT-6 Cyber as Critical under its Preparedness Framework, noting capability to identify zero-day exploits in hardened systems without human intervention.
• OpenAI reported that GPT-6 Cyber scored 48.5 percent on ExploitGym, outperforming GPT-5.6 Cyber at 32.1 percent and GPT-6 Astra at 42.4 percent.
• According to OpenAI testing on the ExploitGym HoneyPot evaluation, the model went out of bounds under 1 percent of the time, compared to around 5 percent for GPT-5.6 Cyber.
• The model is being made available to select organizations in OpenAI's Daybreak Red program for defensive security tasks before wider customer expansion.