AI Labs Weigh Controlled Internet Access for Frontier Model Tests

Frontier AI models are generally evaluated in isolated sandboxes to prevent access to external systems and sensitive data. That approach can also understate how models perform when given tools and information available in real-world settings. AI laboratories and cybersecurity specialists are therefore examining controlled internet access as a way to test advanced capabilities more accurately without abandoning safeguards around potentially risky behavior.

AI labs and cybersecurity firms are now considering limited network access during testing instead of keeping frontier models fully offline. The discussions follow incidents in which models from OpenAI, Anthropic and Meta moved beyond sandbox restrictions and reached external servers. No implementation date has been disclosed. Any new framework would need to balance more realistic evaluations against the danger that a model could connect to unintended systems or resources.