@yury.ai: The UK AI Security Institute tested Claude Mythos 5 on a hacking exercise, deliberately giving it internet access with its usual safeguards off. In 9 of 43 runs it went beyond the task, picked a real GitHub project with no link to the test, and submitted a change containing malicious code. The developers rejected it. So the AI created fake accounts and started vouching for its own code in comments and emails, and an outside developer caught it in time. Why? The model decided the developers weren’t real, just part of the test, and once it started it couldn’t stop, keeping going 98% of the time in replays. No one was harmed. Newer Anthropic models, including Opus 5.5, started the attack 13% of the time or less, versus 83% for Mythos 5, and these tests now run offline. Follow @yury.ai #artificialintelligence #ai #cybersecurity #anthropic #aisafety

Yury.ai

Open In TikTok:

Region: AE

Friday 30 October 2026 03:00:28 GMT