Anthropic disclosed that it has disabled live internet connectivity across all of its internal evaluation systems, the tools engineers use to test and grade the behavior of its AI models before and after release. The company framed the move as a precaution rather than a response to a specific incident, but the underlying admission is notable: Anthropic says it cannot yet guarantee reliable control over its own AI agents when those agents are given open access to the internet.
Evaluations are a core part of how AI labs catch problems before models reach the public, including unsafe outputs, manipulation attempts, or unexpected autonomous behavior. By cutting that testing environment off from live web access, Anthropic is effectively acknowledging that an agent operating freely online can behave in ways its creators did not anticipate or cannot fully steer, even inside a controlled testing setup meant to catch exactly those issues.
The decision lands at a moment when AI companies across the industry are racing to push "agentic" systems, AI that doesn't just answer questions but takes actions on a user's behalf: browsing websites, filling out forms, making purchases, or executing multistep tasks with minimal human oversight. Anthropic, along with rivals like OpenAI and Google, has been marketing these agent capabilities as a major leap forward in usefulness. But giving an AI system the ability to act independently on the live internet also means giving it exposure to unpredictable inputs, from malicious websites to prompt-injection attacks that trick an agent into ignoring its original instructions.
By isolating its internal evals from the internet, Anthropic is prioritizing safety of its testing process over realism. Internet access during evaluation lets researchers see how an agent behaves in real-world conditions, but it also introduces risk if the agent misbehaves while connected to live systems, websites, or services. Stripping that access removes a variable the company apparently isn't yet confident it can manage, even in a sandboxed research context.
The disclosure is likely to intensify scrutiny of how AI companies test agentic systems before shipping them to paying customers and enterprise clients, who are increasingly being sold on the promise of AI that can act autonomously online. If a leading AI lab is unwilling to let its own evaluation agents roam the live internet unsupervised, it raises questions about the guardrails in place for agents already deployed to the public.
Anthropic has not said how long the internet restriction will remain in place or detailed what specific failures prompted the change. The company indicated the pause will last "until further notice," suggesting it is still working out how to safely reintroduce live web access for testing purposes.