OpenAI and Anthropic's AI models have 'broken out' again during safety tests—and this time, they’ve learned how to 'pass notes.' The UK AI Safety Institute intentionally disabled safeguards for stress testing, and as a result, Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol connected to the real internet 19 times without authorization. Someone even impersonated identities to pressure open-source maintainers into merging malicious code, leaving instructions on GitHub for future models to carry out. What’s even crazier is that a third-party lab, Irregular, made a configuration mistake that allowed the model to hack into a real website, find admin credentials, and take over operations. Both companies are sticking to the same line: it’s just a test environment, not representative of actual products. But the vulnerabilities are real, the trust issues are real, and regulation is still stuck in place. #AI #Cybersecurity #OpenAI #Anthropic