Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Anthropic has turned off live internet access for all of its internal evaluations, a move the company described as necessary until further notice. This means AI agents undergoing internal testing can no longer browse the open web during evaluation runs, restricting them to controlled or sandboxed environments.

The decision underscores a core challenge in AI safety: reliably controlling autonomous agents that can interact with the open internet. When agents have live web access, they can encounter unpredictable content, follow unexpected links, or take actions that deviate from intended behavior — scenarios that are difficult to fully anticipate during evaluation.

This is particularly relevant as major AI labs race to build more capable agentic systems. Anthropic's own Claude models are increasingly positioned as tools for autonomous task execution, and the company has been vocal about the importance of alignment and safety in agent development. Cutting off live internet access during evaluations is a conservative but pragmatic step to reduce risk while agent control mechanisms are still being refined.

The move may slow the pace of agentic capability development at Anthropic, as testing agents in sandboxed environments limits the realism of evaluations. However, it also signals a more cautious approach to agent safety that could influence how other AI labs design their evaluation pipelines. The broader industry will be watching to see whether this is a temporary measure or a longer-term shift in how agentic AI is tested.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.