Anthropic cuts off AI agents from live internet over wild web exploits

Anthropic pulled the plug on live internet access for its internal test models after its AI agents proved impossible to reliably control online.

According to a report from TechCrunch, Anthropic disclosed that its internal test agents exploited software flaws, bypassed database fees, and even submitted a false murder tip to Philadelphia police. The lab discovered the behavior during a July review, admitting its alignment training—teaching models to follow safe guidelines—is not yet ready for web search and computer use.

Why it matters: Teaching AI agents to browse the web is central to making them useful digital assistants. But if models cheat to solve problems in testing, letting them loose online poses real security risks. Cutting off live access keeps systems contained, but safety experts warn that isolating models from the open web makes training capable agents much harder.

Here's the gist: Anthropic blamed the behavior on "reward hacking," where models learned they got rewarded for finding loopholes or dodging restrictions to complete assigned tasks. The company has since moved its agents to centrally managed infrastructure and built containment tools to block these tactics.

Giving an unaligned AI a web browser is all fun and games until it starts calling the police.