- Published on
Claude Code Default Permissions Switch to Auto Mode for Pro, Max, and Team Users Starting August 14
- Authors
- Name
- DeepHorus
On August 8, Anthropic announced that starting from August 14, for Pro, Max, and Team subscribers, the default permission mode of Claude Code will be adjusted to automatic mode. This mode automatically intercepts dangerous operations by evaluating tool calls and Shell commands in real time.
The core of automatic mode is an independent AI classifier that performs risk assessment for each operation. Safe operations are allowed through, while destructive, irreversible, or unauthorized behaviors are automatically blocked. Compared to traditional manual approval, this mode aims to improve response speed and security.
Currently, platforms such as Claude Enterprise, Claude API, and AWS remain optional, and users can choose whether to enable automatic mode. Anthropic plans to promote this adjustment on more platforms in the coming month, set it as the default option, and eliminate the corresponding overhead.
During the testing phase, Anthropic invited 1,053 paid testers to participate. The results showed that the traditional manual approval mode identified only 13.6% of dangerous commands, while the automatic mode was able to identify 89% of dangerous commands, a significant improvement in accuracy.
Trajectory Labs conducted prompt injection attack tests, designing a total of 72 scenarios, each tested 10 times, for a total of 720 attack attempts, covering Claude Code v2.1.205 and Codex v0.144.5. In automatic mode, Claude Fable 5, Opus 5, and Sonnet 5 all blocked attacks, with zero successes; while GPT-5.6 Sol had a success rate of 5.83% under Codex's automatic review mode, and as high as 19.03% under full access mode.
This adjustment means that the default security level of Claude Code has been significantly improved, effectively resisting attacks such as prompt injection. For developers, automatic mode can ensure operational safety while reducing manual intervention, improving development efficiency.
Anthropic also pointed out that the bypassPermissions mode without additional protection has an average attack success rate of 0.09%, while the new Auto-review version released by OpenAI last week may change this result. In the future, as automatic mode is fully promoted across platforms, the security standards of AI programming tools may usher in a new landscape.
