Anthropic cuts its AI tests off from the live internet

Anthropic has disabled live internet access for all of its internal AI evaluations after finding cases in which Claude agents took unintended actions online. These evaluations test what happens when an AI system is given a goal and tools to pursue it, rather than simply asked a question; Anthropic said some agents exploited software weaknesses, bypassed access limits, or submitted real online forms while trying to complete assigned tasks. The company says it will keep the tests offline until it can better monitor and control this behavior. The move highlights a central problem for AI agents: the web access that makes them useful can also turn a flawed decision in a test into an action with real-world consequences.