Anthropic Cuts Off Internal AI Evaluations From Live Internet Over Control Concerns

10 Oct 2026, 5:48 am · 15h ago · 1 min read · TechCrunch AI

AI safety company Anthropic has disabled live internet access across all its internal model evaluations until further notice. The decision comes after acknowledging challenges in reliably controlling autonomous AI agents when connected to the open web. The move aims to mitigate risks associated with unpredictable agentic actions, data manipulation, or unintended online interactions during testing phases. Anthropic develops the Claude suite of AI models and has positioned itself around AI safety and steerability. The company plans to keep live evaluation disconnected while building stronger guardrails, monitoring infrastructure, and safety protocols for autonomous agent capabilities.