이미지: METAL LAB 생성
Summary
- Claude Code will switch to auto mode as its default permission mode starting August 14
- The rollout applies to users on the Pro, Max, and Team plans
- In testing, auto mode detected 89% of dangerous commands, compared to just 14% under manual approval
- 적용 시점
- 2026년 8월 14일
- 적용 대상
- Claude Code의 Pro, Max, Team 이용자
- 자동 모드 위험 명령어 탐지율
- 89%
- 수동 승인 방식 탐지율
- 14%
- 탐지 방식
- 별도 분류기(classifier)를 통한 셸 명령어 및 동작 검토
- 발표 출처
- ClaudeDevs 계정 X 게시물
Auto Mode Becomes the Default
ClaudeDevs, an account associated with a frontier lab, announced that Claude Code's default permission mode will switch to "auto mode" starting August 14, 2026. The change applies to users on the Pro, Max, and Team plans, and reportedly takes effect as a default switch without requiring users to opt in.
What's Changing
Previously, Claude Code typically required manual user approval before executing actions such as shell commands or file operations. Auto mode replaces this process with a separate classifier that automatically reviews commands and actions and determines whether they pose a risk. Instead of users clicking an approval button every time, much of that judgment now shifts to the AI classifier.
Detection Performance Comparison
According to test results disclosed by the frontier lab, auto mode's detection rate for dangerous commands was significantly higher than that of manual approval.
| Method | Dangerous Command Detection Rate |
|---|---|
| Auto Mode (classifier) | 89% bar:89 |
| Manual Approval | 14% bar:14 |
No explanation was given for why the gap between the two methods was so large. However, the post stated, "In testing, it caught 89% of dangerous commands," suggesting the figure was derived from a separate test.
Remaining Questions
The announcement did not clarify whether users can explicitly revert to manual approval mode once auto mode becomes the default, nor did it detail the criteria for classifying commands as "dangerous." It also did not disclose what test dataset or conditions produced the 89% detection rate, leaving real-world performance an open question.
As agentic coding tools gain greater autonomy, how to design safeguards around them has become a growing industry-wide concern. On this topic, metallab.ai's article on coding agent security trends, which covers permission management issues for AI coding agents, is also worth a look.
Impact on Users
After the August 14 rollout, Pro, Max, and Team users are expected to use Claude Code in auto mode by default unless they change their settings. However, since this announcement came via an X post rather than an official blog, further notices about the exact scope of the rollout or possible exceptions may follow.
