Anthropic severs live internet from all internal AI evals indefinitely

Anthropic, the AI lab behind the Claude models, has turned off live internet access for all its internal evaluations until further notice. The announcement, reported by TechCrunch, is a single sentence with no explanation. For a company that has built its brand on safety-first AI, the move is a striking admission that something about live internet testing is not working.

Internal evaluations are the tests Anthropic runs to measure how its models perform on real-world tasks. When those evals have live internet access, the AI can browse websites, pull up-to-date information, and interact with online services. That is exactly the kind of behavior that AI agents are supposed to handle. Cutting off that access means Anthropic can no longer verify that its models behave safely and reliably in live conditions. The "until further notice" qualifier suggests this is not a quick fix.

To understand why this matters, consider the state of AI agents. Over the past year, every major lab, including OpenAI, Google, Microsoft, and Anthropic, has been racing to ship systems that can act autonomously: book travel, manage email, make purchases, and operate software. These agents are designed to use the internet as their playground. But with that power comes risk. An agent could be manipulated by malicious websites, leak sensitive data, or take actions with unintended consequences. The industry has been grappling with how to test these systems safely.

Anthropic's decision to sever live internet from all internal evals is a clear signal that the company does not yet trust its agents to roam free. It is one thing to run a model in a sandbox with curated data; it is another to let it loose on the open web, where it can encounter adversarial content, changing information, and unpredictable APIs. By pulling the plug, Anthropic is effectively saying that its current evaluation infrastructure cannot guarantee safety in live conditions.

This is not the first time Anthropic has taken a cautious stance. The company has long positioned itself as the safety-conscious lab, with a charter focused on beneficial AI and a track record of releasing models with careful guardrails. But this move goes beyond a single model release. It affects the entire internal evaluation pipeline, which means every new model and agent update will be tested without live internet until further notice. That could slow down Anthropic's development cycle, giving rivals a window to push ahead, or it could give Anthropic a more robust safety foundation in the long run.

For executives and founders building AI products, the implications are immediate. If a leading lab like Anthropic cannot reliably control its AI agents in live internet environments, then your own team's agents are likely facing similar challenges. The move is a reminder that offline testing, sandboxing, and human oversight are not optional extras; they are core requirements for any autonomous system. It also raises questions about how you evaluate your own models. Are you testing with live internet? If so, what safeguards do you have in place? This is especially critical for companies that are already deploying agentic features in production, where a single failure can erode customer trust and invite regulatory scrutiny.

Regulators are watching. Governments on both sides of the Atlantic have been drafting rules for AI, with a particular focus on high-risk applications. The EU AI Act, for example, imposes strict requirements on systems that interact with the public. An incident involving an AI agent that goes rogue online could accelerate regulatory action. Anthropic's move may be a preemptive step to avoid such an incident, but it also signals to regulators that the technology is not yet mature. The absence of a clear timeline from Anthropic only adds to the uncertainty.

The strategic stakes are clear. For Anthropic, the decision to cut live internet access is a bet that safety is worth the cost of slower progress. For everyone else, it is a wake-up call: the gap between a demo and a dependable agent is still wide. The companies that figure out how to test and deploy agents safely, without cutting off the internet entirely, will have a massive advantage. Until then, the rest of us should take note: if Anthropic is pulling the plug, you should be checking your own connections.