Anthropic Pauses Internal AI Evaluations to Address Agent Control Concerns
Anthropic has temporarily disconnected its internal evaluation systems from the live internet. The company announced it had "turned off live internet access" for all internal evaluations until further notice, a move prompted by challenges in maintaining reliable control over AI agents navigating real-world web conditions.
The decision reflects broader concerns in the AI industry about agentic systems—AI models designed to take actions autonomously, such as browsing the web, filling forms, or interacting with external services. As these systems become more capable, ensuring they behave predictably and safely, especially when accessing unconstrained online environments, has proven technically difficult.
The suspension means Anthropic's teams will rely on simulated or sandboxed testing environments rather than live web interactions for now. While evaluation pipelines typically use controlled benchmarks to measure model capabilities, some assessments require real-world conditions to gauge performance accurately. Pausing live internet access could slow certain research and development processes but also signals a cautious approach to safety as the company works to improve agent control mechanisms.
Anthropic has not specified a timeline for restoring live internet access to its internal evaluations.