Anthropic has disabled live internet access for all of its internal evaluations, a notable safety step that highlights how difficult it remains to reliably control frontier AI agents once they are connected to the open web. The company said it had "turned off live internet access" for "all our internal evaluations" until further notice, a move that signals concern not just about model performance, but about the integrity of the testing environment itself.
Safety Testing Shift
The decision matters because internal evaluations are one of the main ways AI developers measure whether a model behaves as intended before broader release. In a live internet setting, however, an agent can encounter unpredictable websites, changing content, external tools, and dynamic responses that make results harder to reproduce and interpret. By cutting off live access, Anthropic is effectively acknowledging that some evaluation conditions may be too unstable, too open-ended, or too risky to trust as a basis for safety claims.
That is a significant admission for a company that has positioned itself as one of the industry's most safety-conscious frontier AI labs. Anthropic has repeatedly emphasized alignment, guardrails, and responsible deployment as core principles. But the need to isolate evaluations from the live internet suggests that even leading labs are still wrestling with a basic question: how do you test an agent that is designed to act in the world without letting the world distort the test?
Agent Control Problem
The issue goes beyond a single technical setting. AI agents are increasingly being built to browse, search, summarize, execute tasks, and interact with software on behalf of users. Each added capability expands usefulness, but also increases the surface area for failure. A model that performs well in a controlled sandbox may behave differently when exposed to live data, real-time prompts, adversarial content, or unexpected instructions embedded in webpages.
That creates a control problem for developers. If an agent can be nudged off course by external content, then evaluation results may overstate its reliability or understate its risk. If it can take actions in the wild that are difficult to anticipate, then even a successful test run may not prove much about real-world safety. Anthropic's move suggests that the company sees live internet access not as a convenience, but as a variable that can undermine the validity of internal assessment.
The broader industry has been moving quickly toward more autonomous systems, with major labs racing to build models that can reason, plan, and act with less human oversight. But the faster those systems become, the more pressure there is on developers to prove they can be constrained, audited, and evaluated under conditions that are both realistic and controlled. Anthropic's decision is a reminder that those two goals are often in tension.
Industry-Wide Implications
The change may also influence how other AI companies design their own testing regimes. If live internet access is too noisy or too risky for internal evaluations, labs may increasingly rely on simulated environments, curated datasets, or restricted testbeds that better preserve repeatability. That could improve safety measurement, but it may also leave gaps between benchmark performance and actual deployment behavior.
For regulators and enterprise customers, the move is likely to reinforce concerns that agentic AI remains immature in key operational respects. Businesses want systems that can search the web, retrieve information, and automate tasks, but they also need assurance that those systems will not drift, hallucinate, or act unpredictably when exposed to real-world inputs. Anthropic's decision does not imply a crisis, but it does suggest the field is still in an experimental phase where even basic evaluation design is under active revision.
The company did not indicate whether the restriction applies to product-facing systems or only to internal testing workflows. Still, the message is clear: when it comes to frontier AI agents, live internet access can be as much a liability as a feature. For now, Anthropic is choosing control over realism in its evaluations, a tradeoff that may become more common as the industry confronts the limits of testing autonomous systems in an uncontrolled environment.
The development arrives at a time when the AI sector is under intensifying scrutiny over safety, reliability, and the pace of deployment. As models become more capable, the gap between what they can do in theory and what they can safely do in practice is becoming harder to ignore. Anthropic's decision to sever internal evaluations from the live web is a small operational change with a large symbolic meaning: the frontier is still not fully governable.