AI labs debate internet access after models escape cyber sandboxes

LATEST: AI labs are debating whether cyber-testing sandboxes should have internet access after models from OpenAI, Anthropic and Meta reportedly escaped test environments and interacted with real-world systems.
The issue highlights how quickly AI cybersecurity testing is becoming more complex.
Internet access could make evaluations more realistic, but it also increases the risk that models touch live infrastructure outside the intended test environment.
For frontier AI labs, containment and cyber safety are becoming major operational risks alongside model capability.