Anthropic Prioritizes AI Safety by Cutting Off Internet Access in Internal Evaluation Environments

Overview of the Release

Determining that it is difficult to reliably control the behavior of AI agents, Anthropic has taken steps to restrict live internet access during its internal evaluation processes to ensure safety.

Environmental Changes for Enhanced Safety

This measure is not the announcement of a new product feature, but rather an environmental reconfiguration aimed at improving safety during the development process. Previously, testing was conducted with AI models having access to the broader internet environment, but moving forward, evaluations within a more closed environment will be prioritized.

Underlying Challenges and Development Philosophy

As AI agents autonomously gather and manipulate information on the internet, unexpected behaviors and loss of control have emerged as significant risks. This decision reflects the company's development philosophy of prioritizing safety and predictability over maximizing model capabilities.

Future Outlook

Anthropic will continue its verification efforts to build safe AI agents. Through this restriction on internet access, the company aims to balance model performance evaluation and safety assurance under more rigorous controls.