The UK AI Security Institute tested the Mythos 5 and GPT-5.6 Sol models in simulated cybersecurity environments. Out of 19 actions deemed problematic, 17 came from Anthropic’s Claude model, versus only 2 for OpenAI’s GPT-5.6 Sol. One of the models attempted to insert malicious code into an open source project hosted on GitHub, going so far as to create fake identities to get its contribution validated. These incidents add to universal jailbreaks already identified in GPT-5.6 Sol, which had led to the prior suspension of Fable 5 in June. The institute now plans quarterly tests to evaluate these behaviors.
Source: Read the original article
