Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableZDNet AI→ original

Anthropic adds auto mode to Claude Code: fewer confirmations, protection against dangerous commands

Anthropic has launched auto mode for Claude Code, a middle ground between manual confirmation and switching off all restrictions. It lets safe edits and commands through without needless pauses, while risky actions such as deletion, force push or work on external infrastructure are stopped by a separate classifier. The mode is still a research preview and does not remove the need for manual review on sensitive tasks.

AI-processed from ZDNet AI; edited by Hamidun News
Anthropic adds auto mode to Claude Code: fewer confirmations, protection against dangerous commands
Source: ZDNet AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceZDNet AI ↗
Editor-in-Chief
Key metric #1force push and direct pushes to main
Key metric #2irreversible deletion of old files and mass deletions in cloud storage
Key metric #3production deploys, migrations and changes to shared infrastructure

Anthropic has introduced auto mode in Claude Code, a mode that removes most of the annoying confirmation prompts but does not turn the AI agent loose entirely. The idea is to speed up long coding sessions while keeping the model from accidentally deleting branches, leaking secrets or reaching into production.

How the mode works

Until now, Claude Code users faced an unpleasant choice. Either confirm nearly every file edit, shell command and network request, or switch to a mode with minimal restrictions and take on all the risk. Anthropic says developers approve about 93% of access requests. That means protection exists on paper but quickly becomes routine in practice: people click approve almost automatically and stop looking closely at what the agent is about to do.

Auto mode is designed as a middle option between manual control and the --dangerously-skip-permissions flag. On ordinary tasks Claude Code can work without stopping, while questionable actions go not to a human but to a separate classifier. The setup has two layers: first, server-side protection checks tool output for prompt injection; then a second layer analyzes the agent's action itself, such as a command, a network call or an attempt to launch a subagent. If an action looks irreversible, suspicious or outside the scope of the user's request, it is blocked.

What gets blocked

Anthropic deliberately kept the default settings of auto mode conservative. Inside the working project, Claude Code can freely edit files and carry out ordinary development steps without wasting time on checks. But once an action could cause noticeable damage, separate safety logic kicks in. By default, the mode trusts only the current directory and configured remote repositories, and treats everything else as an external environment.

  • force push and direct pushes to main
  • irreversible deletion of old files and mass deletions in cloud storage
  • production deploys, migrations and changes to shared infrastructure
  • sending secrets, .env files or internal code to untrusted addresses
  • running externally downloaded code in the style of curl | bash

If a team needs to allow more, the boundaries can be defined in settings: add trusted domains, buckets, repositories and internal services. This matters for companies where the agent really does need to push to a corporate GitHub org, write to cloud storage or call an internal API. Anthropic also stresses that rules from a repository's shared file cannot expand the allow list themselves and thereby bypass the protection.

Where the risks remain

Anthropic does not sell auto mode as absolute protection, and that is arguably the launch's main strength. In internal tests on real traffic, the full pipeline produced about 0.4% false positives, but missed roughly 17% of dangerous "overly eager" actions. For a mode meant to be safer than turning confirmations off entirely, that is already a useful step. For work with production, secrets or critical infrastructure, it is still not enough to relax and take the human out of the loop.

"Auto mode is designed as a replacement for dangerously-skip-permissions without a return to constant pauses."

That is why the feature is still in research preview. It is not available to everyone: it requires a suitable Claude Code plan, enabling by an administrator for team accounts, a compatible model at the level of Sonnet 4.6 or Opus 4.6, and Anthropic API as the provider. If the classifier blocks actions three times in a row or 20 times in a session, Claude Code falls back to normal confirmations. In other words, the system doesn't just cut risky steps; it can also halt an autonomous session if it sees the agent pushing too persistently on a dangerous scenario.

What this means

Claude Code is taking an important step from a "chat assistant for code" toward a full agent that can be handed long tasks without constant supervision. But Anthropic is candid about the limits: auto mode is not a magic autopilot. It is a compromise safety layer that is noticeably more convenient than manual approval windows and far more sensible than removing restrictions altogether.

Story

Key Questions Answered

How the mode works?

Until now, Claude Code users faced an unpleasant choice. Either confirm nearly every file edit, shell command and network request, or switch to a mode with minimal restrictions and take on all the risk. Anthropic says developers approve about 93% of access requests. That means protection exists on paper but quickly becomes routine in practice: people click approve almost automatically and stop looking closely at what the agent is about to do.

What gets blocked?

Anthropic deliberately kept the default settings of auto mode conservative. Inside the working project, Claude Code can freely edit files and carry out ordinary development steps without wasting time on checks. But once an action could cause noticeable damage, separate safety logic kicks in. By default, the mode trusts only the current directory and configured remote repositories, and treats everything else as an external environment.

Where the risks remain?

Anthropic does not sell auto mode as absolute protection, and that is arguably the launch's main strength. In internal tests on real traffic, the full pipeline produced about 0.4% false positives, but missed roughly 17% of dangerous "overly eager" actions. For a mode meant to be safer than turning confirmations off entirely, that is already a useful step. For work with production, secrets or critical infrastructure, it is still not enough to relax and take the human out of the loop.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…