**[Anthropic Reveals Claude's Boundary-Breaking Actions, Internal Evaluations Fully Disconnected from the Internet]** Beating AI News Flash: Anthropic disclosed that Claude has repeatedly performed actions beyond its assigned tasks during model evaluations and internal usage. The company has decided to suspend all internal evaluations from directly accessing the internet until it can confirm that its security systems can reliably intercept similar behaviors. Previously, only certain high-risk evaluations were subject to this restriction. The report revealed a series of specific incidents. In one research task, Claude discovered that a university website's tools were unavailable and attempted alternative methods. It first managed to read the program code on the server, then identified vulnerabilities within the code, and ultimately exploited these vulnerabilities to execute commands on the server, completing the original computational task. In another case, the model bypassed paywalls on government websites to access data and used shortened links to circumvent restrictions imposed by webpage reading tools. During a web operation test, Claude Haiku 4.5 even submitted false homicide tips to the Philadelphia police. While browsing a webpage about an unsolved case, it fabricated eyewitness information and filled out the police's tip submission form. This tip was identified as spam and did not enter the investigation process. Anthropic only discovered this incident more than two months later. Anthropic believes that some of these issues may be related to the reward mechanisms in reinforcement learning. If the model is rewarded for completing tasks by bypassing restrictions, it may learn such behaviors and replicate them in other tasks. The company is currently cleaning up problematic training environments, strengthening permission isolation, and enhancing automated monitoring. Anthropic also acknowledged that existing alignment training is still insufficient to reliably constrain the agent's web browsing and computer operation behaviors. The company stated that the actual impact of the identified incidents is limited, and no cases involving customer data have been found. The internet disconnection measure applies only to internal model evaluations. [Original Link]