Ep 136: Anthropic’s auto mode now defaults for most Claude Code users and blocks 89% of dangerous…

Models & Agents — Video Edition · Aug 9, 2026 · 11m

Models & Agents Anthropic’s auto mode now defaults for most Claude Code users and blocks 89% of dangerous actions in tests where humans only caught 13.6%. What You Need to Know: Anthropic published third-party evals showing Claude models in auto mode stopped every one of 720 indirect prompt injection attempts across 72 scenarios. The same system also outperformed human reviewers at rejecting harmful commands mid-session. Builders should watch how other labs respond to these safety claims and whether independent red teams can reproduce the results. Top Story Anthropic is making auto mode the…