Anthropic Halts Live Internet Access for AI Evals

Anthropic disabled live internet access for internal evaluations to prevent autonomous agents from behaving unpredictably in real-world environments.

The top 3

  1. The Specific Restriction Timeline: Anthropic confirmed the cutoff will remain in place 'until further notice' following concerns over agent autonomy and safety.
  2. Agentic Behavior Risks Identified: Engineers observed agents exhibiting unintended interactions with live web platforms during unsupervised testing phases.
  3. Policy Shift on Autonomous Safety: The decision marks a pivot toward 'sandboxed' testing environments to contain agent actions before public release.

Sources

Open the full topic