Anthropic restricts internet access for internal AI agent evaluations

Anthropic has announced a significant change to its safety testing protocols, confirming that it has disabled live internet access for all internal evaluations of its AI agents. This decision follows concerns regarding the company's ability to reliably control autonomous agents when they are connected to the live web. By cutting off external connectivity during the evaluation phase, Anthropic aims to mitigate risks associated with unpredictable agent behavior and potential security vulnerabilities. The company stated that this restriction will remain in place until further notice as they refine their safety frameworks. This move highlights the ongoing industry-wide struggle to balance the capabilities of autonomous AI systems with the necessity of maintaining rigorous safety and control standards. As Anthropic continues to develop its agentic models, the focus remains on ensuring that these systems operate within secure, predictable boundaries before they are deployed in broader, real-world environments.