Anthropic Suspends Internet Access for AI Agents in Internal Testing

Anthropic, one of the leading companies in the development of artificial intelligence systems, has announced an unusual measure: suspending internet access for all internal evaluations of its AI models. The decision comes as the company admits it cannot reliably control the behavior of its autonomous agents when they operate with unrestricted network access, according to TechCrunch.

What the change means

According to the cited source, the company informed staff internally that it had "turned off live internet access" for all evaluation tests conducted on its AI models. The measure remains in effect until further notice, with no clear timeline yet given for restoring unrestricted access.

Internal evaluations are an essential step in the development of any AI system, as they allow engineers to test how models respond in a variety of scenarios, including situations where agents must browse the internet to complete complex tasks. By cutting off this access, Anthropic appears to be trying to limit the risks associated with unforeseen or potentially problematic actions by its systems.

Concerns over the autonomy of AI agents

Anthropic's decision highlights a broader issue facing the artificial intelligence industry: the difficulty of guaranteeing full control over autonomous AI agents, particularly when they can interact freely with the online environment. AI agents are designed to carry out complex tasks, sometimes with minimal human oversight, which increases the risk of unexpected or hard-to-predict behavior.

This public acknowledgment of the technology's current limitations comes at a time when companies in the field are under pressure to demonstrate both technological progress and robust safety measures. Anthropic has consistently positioned itself as a company focused on AI safety, and this decision appears to reaffirm that commitment to a cautious approach, even if it means temporarily restricting its own testing processes.

It remains unclear, for now, whether this restriction will also affect the company's commercial products aimed at end users, or whether it is limited strictly to internal evaluation and development processes.