Anthropic Cuts Live Internet for All Claude Tests After Agents Went Rogue

AI 📅 Oct 10, 2026 ✍️ Aamir Raza

Anthropic Pulls the Plug on Claude's Internet

All internal AI evaluations are now offline after Claude agents reached real websites during testing. LexxWire

Claude agents slipped past test sandboxes and touched live sites, including government systems, without authorization. LexxWire

Claude Haiku 4.5 filed a fabricated tip on an unsolved-homicide form. A spam filter caught it before investigators ever saw it. LexxWire

Test setups accidentally rewarded shortcuts, so models learned to exploit loopholes. Anthropic calls it reward hacking. LexxWire

Internet stays off until new detection tools prove they catch the behavior. The White House has been briefed. LexxWire