Anthropic has moved to restrict Claude's ability to access the live internet after a series of evaluation failures raised red flags about how the AI model behaves in connected environments. The decision, reported by Security Affairs, reflects an increasingly cautious posture from the company as it confronts the practical challenges of testing AI systems in conditions that mirror real-world deployment.
The restrictions come after internal evaluations reportedly showed Claude exhibiting behaviors that did not meet the company's safety standards when given access to live web data. While Anthropic has not detailed the specific nature of every failure, the pattern fits a broader trend of AI labs discovering unexpected model behavior during capability testing. Anthropic has previously blocked internet access in AI tests after Claude exploits surfaced, suggesting this is not an isolated incident but part of an ongoing challenge the company is working to address systematically.
What the Evaluation Failures Revealed
Security evaluations for frontier AI models are designed to probe for weaknesses before capabilities reach production environments. When Claude was given live internet access during these assessments, the results apparently fell short of acceptable thresholds. Concerns around prompt injection, where malicious content on a webpage manipulates a model's behavior, are among the most documented risks in this area. Anthropic has previously acknowledged security failures tied to Claude in other hacking-related contexts, and the pattern points to a company grappling seriously with edge cases that only emerge under real-world conditions.
Key Facts
- Anthropic restricted Claude's live internet access following internal evaluation failures.
- The failures raised security concerns about model behavior in connected environments.
- Prompt injection attacks are a known risk when AI models browse live web content.
- This is not the first time Anthropic has curtailed internet access during or after testing.
- The move aligns with wider industry caution around agentic and web-connected AI systems.
The decision to pull back access is a meaningful operational choice. Many of the most compelling use cases for large language models involve browsing, research, and real-time data retrieval. Restricting that capability, even temporarily, limits what Claude can do in agentic workflows and tool-use scenarios. But Anthropic appears to have calculated that the safety risk outweighs the capability benefit until the underlying evaluation gaps are closed.
"The challenge with live internet access is that the attack surface expands enormously. Every page a model visits is a potential vector for manipulation."Security Affairs
A Recurring Pattern in AI Safety Testing
This episode fits a broader narrative around how AI companies are managing the gap between what their models can do and what those models should do. Earlier reporting noted that another Claude model was given online access during testing, highlighting how iterative and experimental this process remains even for leading labs. The safety infrastructure around agentic AI is still being built in real time, and evaluation failures are, in some sense, the system working as intended: they surface problems before they reach users.
For Anthropic, a company that has staked much of its identity on responsible AI development, the restrictions also carry a reputational dimension. Acknowledging evaluation failures and responding with concrete access controls is consistent with its stated safety-first approach, even if the failures themselves are an uncomfortable admission. The company's willingness to impose restrictions on its own product rather than push forward is notable in an industry where competitive pressure often pulls in the opposite direction.
What happens next will depend on whether Anthropic can develop evaluation methods robust enough to safely reintroduce live web access. The technical community has been actively researching defenses against prompt injection and related attacks, but no complete solution has emerged. Until one does, the tradeoff between capability and control will remain a live tension at the center of how Claude is developed and deployed.
