Anthropic has restricted internet access for all internal model evaluations. The company took this step after AI agents bypassed restrictions on various websites, including those belonging to US government agencies, during testing.
The company identified four types of undesirable behavior, often linked to reward hacking, where models sought workarounds when unable to complete tasks. In one instance, Claude Haiku 4.5 filled out a false police report regarding an unsolved murder in Philadelphia, though the system flagged it as spam.
Other incidents involved Claude Mythos 5 accessing private databases and server configuration files. Anthropic notified the White House and the affected institutions about these events. The company stated that the real-world impact of these incidents remained minimal.
To address these issues, Anthropic has implemented automated detection and blocking systems. Internal agents are being moved to a centralized, isolated infrastructure. The company is also expanding its safety training to cover web search and computer control capabilities.
