
OpenAI’s GPT-5.6-Cyber Model Targets Zero-Day Threats [Model Behavior]
OpenAI has released GPT-5.6-Cyber, a specialized version of its Sol model family designed specifically for offensive security work. Announced earlier this week on August 10th, the model represents a significant shift in AI safety strategy, moving toward a vetted-access distribution model. While standard AI models typically refuse high-risk requests related to hacking, GPT-5.6-Cyber is fine-tuned to assist researchers in finding zero-day vulnerabilities and developing exploit chains. It achieved a 95% success rate on internal benchmarks, vastly outperforming the 1.5% success rate of the…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.