OpenAI Previews GPT-5.6 Sol for Cybersecurity Tasks

Summary

OpenAI is previewing GPT-5.6 Sol to approved partners, describing it as its strongest cybersecurity model. The company reports vulnerability-finding and patch-development strengths, alongside real-time safeguards and automated red-teaming.

Key points

OpenAI says Sol matched competing systems on ExploitBench while using roughly one-third the output tokens.
In tests on Chromium and Firefox, Sol reportedly identified bugs and basic exploitation primitives but did not independently build a full-chain exploit.
OpenAI says the model is better suited to defensive tasks such as vulnerability identification and patch development than end-to-end cyberattacks.
The GPT-5.6 series uses real-time classifiers for cybersecurity and biology inputs; flagged outputs may be paused for secondary review.
OpenAI says it used more than 700,000 A100-equivalent GPU hours for automated red-teaming focused on universal jailbreaks.
Access is currently limited to approved partners through the API and Codex; broader availability is planned in the coming weeks.

Article Details

Event Type: Limited preview launch of OpenAI’s GPT-5.6 model lineup, including Sol, which OpenAI describes as its most capable model for cybersecurity.
Impact: Sol demonstrated vulnerability identification and patch-development capabilities in evaluations on Chromium and Firefox codebases, isolating bugs and basic exploitation primitives but not independently constructing a full-chain exploit. OpenAI says the models emphasize defensive security and use refusal training, real-time classifiers, and secondary review to manage dual-use risks. Access is currently limited to approved partners through the API and Codex, with broader availability planned in the coming weeks.