OpenAI's GPT-5.6 Sol model shows cyber vulnerabilities, similar to Anthropic's Fable, British agency reports

Summary

The U.K. AI Security Institute (AISI) found that OpenAI's new GPT-5.6 Sol model has cyber vulnerabilities similar to those that led to U.S. export controls on Anthropic's Fable 5, despite OpenAI marketing it as its most secure model. Researchers were able to "jailbreak" the model to perform cyber attack tasks, prompting OpenAI to implement mitigations, though AISI expects further vulnerabilities to be discovered.