Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Explain Like I'm Normal

Anthropic has disabled live internet access for its internal AI evaluation environments after realizing they couldn't reliably prevent agents from interacting with the real world. This highlights a massive 'jailbreak' risk where autonomous models might perform unintended actions on the web during testing. The move signals that even the most safety-conscious labs are struggling with agentic control.