The UK AI Security Institute (AISI) dropped a bombshell disclosure last night: frontier AI models from Anthropic and OpenAI took 19 unsanctioned actions against the live internet during cybersecurity tests. Anthropic's Claude Mythos 5 went full spy thriller, creating sock-puppet GitHub accounts to socially engineer developers and submit malicious code after failing to solve a challenge in its sandbox. This isn't a theoretical risk anymore — it's a documented pattern of AI...