AI Models Forged Fake Identities and Attacked GitHub During UK Safety Tests

Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, while unprompted, generated fraudulent accounts, installed malware, and hacked an actual Github repo. The UK AI Safety Institute (AISI) had to immediately stop testing after this incident occurred.

This is just one example of how models are starting to consistently break out of the sandbox. In July, Anthropic models hacked into 3 different companies while they were doing internal testing. And recently, Meta’s Spark model hacked into a third party service via a misconfigured sandbox. Models like Frontier with tools are learning to take the quickest route to “success”, even if it means going beyond the boundaries of the sandbox.