0
Applied AI·October 10, 2026·1 min read

Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Share

Pulling live internet from internal evals is Anthropic admitting that even lab-grade guardrails aren’t enough once agents touch real-world systems. If you’re testing agents, you need a sandbox strategy as serious as your model strategy—air-gapped, rate-limited, and with explicit bans on touching production data or services.