GPT‑6 Cyber scores 48.5% on ExploitGym, up from 32.1%

GPT‑6 Cyber builds directly on GPT‑5.6 Cyber, and OpenAI frames the gap between the two as a step change rather than an incremental gain. On ExploitGym, which tests whether a model can exploit a specified vulnerability, the score rises from 32.1% to 48.5%. OpenAI also says the model outperforms GPT‑6 Astra on most cyber tasks even when Astra does not refuse the request, and it now classifies GPT‑6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT‑5.6 Cyber.

The alignment results are presented as a direct response to a known failure mode: security agents that go outside the scope of their assigned task. In the ExploitGym Honeypot evaluation, which was informed by the Hugging Face incident, the model must infer that exploiting easy-to-attack supporting services is out of scope. GPT‑6 Cyber avoided those services in 99.93% of cases without deployment safety prompts, up from 94.49% for GPT‑5.6 Cyber, and 100% of the time with those prompts.

OpenAI ties the release to a narrow window for using frontier models to secure software, and says its goal is to dramatically expand Daybreak to organizations of every size, independent researchers, and open-source maintainers. The same release pairs stronger capabilities with safeguards aimed at misuse, including cyber-specific misalignment monitoring and a $1 billion Daybreak for Frontline Defenders grant program for resource-constrained defenders such as government organizations and essential service providers.