Anthropic adds AI-judged auto mode to Claude Code
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic launched "auto mode" for Claude Code in research preview, letting its AI judge which actions are safe to execute automatically while blocking risky ones
- The safety layer reviews each action before it runs, screening for risky behavior the user didn't request and for signs of prompt injection — malicious instructions hidden in processed content
- Auto mode extends Claude Code's existing "dangerously-skip-permissions" command, which hands all decisions to the AI, by adding a safety check on top
- Anthropic has not disclosed the specific criteria its safety layer uses to distinguish safe actions from risky ones, a gap developers will likely want filled before adopting it widely
- The feature is limited to Claude Sonnet 4.6 and Opus 4.6 and rolls out to Enterprise and API users in the coming days
- Anthropic recommends running auto mode in "isolated environments" — sandboxed setups kept separate from production systems to limit damage if something goes wrong
Why it matters: For developers, auto mode replaces a binary trade-off between constant approval prompts and fully hands-off execution, but Anthropic's opacity about the safety layer's decision criteria — combined with its own advice to sandbox the feature — signals meaningful residual risk the company itself acknowledges. Enterprise and API users get first access in the coming days, restricted to the two newest Claude models, which means broader developer adoption is gated on both model access and whatever transparency Anthropic eventually provides about its filtering logic.
