Every Claude Code user knows the drill: approve this file write, approve that bash command, approve, approve, approve. Anthropic just shipped a fix.
Auto mode is a new permission level that sits between Claude Code's default (ask before everything) and the ominously named --dangerously-skip-permissions flag (ask before nothing). Before each action, an AI classifier evaluates whether the operation is risky. Safe tasks like writing a new component file proceed without interruption. Potentially dangerous ones, like mass-deleting files, exfiltrating data, or running suspicious scripts, get blocked automatically. If Claude keeps trying to do something the classifier flags, it escalates to a manual approval prompt.
The classifier also watches for prompt injection, where malicious instructions are hidden inside code comments, documentation, or log output to trick the model into doing something harmful. That's a real and growing attack vector for coding agents, so it's good to see Anthropic addressing it directly.
The Three Permission Levels Now
- Default mode: Every file write and bash command needs your approval. Safe but slow.
- Auto mode: The classifier screens each action. Safe operations run automatically, risky ones get blocked or escalated.
--dangerously-skip-permissions: No guardrails at all. The name is the warning label.
Availability
Auto mode is available now as a research preview for Claude Teams users. Enterprise and API access is coming "in the coming days," according to Anthropic. The company recommends running it in sandboxed environments rather than anywhere near production systems.
Anthropic is careful to note that auto mode "reduces risk compared to --dangerously-skip-permissions but doesn't eliminate it entirely." That's an honest framing. Coding agents that can write files and execute shell commands are inherently risky, and no classifier is perfect.
For anyone doing extended coding sessions where the constant permission prompts break flow, this is a practical improvement. It won't replace careful review of what your AI assistant is doing, but it should cut the friction on routine operations significantly.