Claude Code 本月14日起默认自动模式:危险操作AI直接拦下 Claude Code Will Default to Auto Mode on Aug 14: AI Blocks Dangerous Commands
Anthropic官宣:从8月14日起,Claude Code的Pro、Max和Team订阅用户会默认启用“自动模式”。这个模式相当于给AI编程助手装了个安全门卫,每次执行命令前先由AI分类器看一眼,安全操作放行,破坏性、不可逆或越权操作直接拦下。安不安全不是玄学,测试数据显示,人工审批只能认出13.6%的危险命令,自动模式能认出89%。对普通用户来说,以后用Claude写代码会更省心,但默认权限的改变也意味着AI的判断越来越重要。企业版、API等平台暂时不变,预计未来一个月内跟进。
Anthropic announced that starting Aug 14, Pro, Max, and Team subscribers will get Claude Code’s new default permission setting: Auto Mode. Think of it as a safety guard for AI coding: before each command or tool call runs, a classifier lets safe actions through and stops destructive, irreversible, or unauthorized operations. In testing, human approval caught only 13.6% of dangerous commands, while Auto Mo
## 背景 以前用Claude Code,开发者要手动确认每条Shell命令,安全是安全,但很打断思路。8月14日起,Pro、Max和Team订阅用户会默认切到“自动模式”。这模式不是AI乱放行,而是由独立的AI分类器实时判断:安全操作直接执行,有风险、不可逆或越权的操作自动拦下来。 ## 关键细节 - 覆盖范围:Pro/Max/Team先改;Enterprise、API、AWS Bedrock、Google Cloud平台、Microsoft Foundry等暂时仍是可选,但Anthropic计划在未来一个月内把这些平台也默认改成自动模式,并取消额外开销。 - 测试数据:Anthropic拉了1053名付费测试者做对比,人工审批只识别出13.6%的危险命令,自动模式识别出89%。 - 提示词注入测试:Trajectory Labs用72个场景、每个场景10次,共720次攻击,覆盖Claude Code v2.1.205和Codex v0.144.5。Claude Fable 5、Opus 5和Sonnet 5在自动模式下全部拦截,0次成功。对比来看,OpenAI的GPT-5.6 Sol在Codex的Auto-review模式下成功率为5.83%,在Full Access模式下为19.03%。Anthropic还提到,没加额外防护的bypassPermissions模式平均攻击成功率为0.09%,而OpenAI上周发布的新Auto-review版本可能改变结果。 ## 影响 这次调整把“安全判断”的责任从人交给AI,本质上是让AI工具在效率和安全之间找新平衡。对个人开发者来说,默认自动模式可以减少确认弹窗,写代码更顺手;但危险操作的拦截精度直接决定用户信任度。企业用户接下来如果也默认切到自动模式,意味着AI厂商要对权限闸门承担更大责任,也说明AI Agent正在从“辅助”走向“半自主”阶段。
## Background Claude Code, Anthropic's AI coding tool, used to require human approval for each shell command. That felt safe but broke developers' flow. Starting Aug 14, Pro, Max, and Team subscribers will instead default to Auto Mode. In this mode, an independent AI classifier does real-time checks: safe operations run automatically, while risky, irreversible, or unauthorized actions are blocked. ## Key details - Coverage: Pro/Max/Team first; Enterprise, API, AWS Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry remain optional for now, but Anthropic plans to make Auto Mode the default there within about one month and remove extra overhead. - Test data: Anthropic asked 1,053 paid testers to compare processes. Human approval detected only 13.6% of dangerous commands, while Auto Mode detected 89%. - Prompt injection testing: Trajectory Labs ran 72 scenarios x 10 tries, for a total of 720 attacks, covering Claude Code v2.1.205 and Codex v0.144.5. Claude Fable 5, Opus 5, and Sonnet 5 in Auto Mode blocked all attacks, with 0 successes. By comparison, GPT-5.6 Sol in Codex's Auto-review mode succeeded 5.83% of the time, and in Full Access mode it succeeded 19.03%. Anthropic also noted that an unprotected bypassPermissions mode had an average attack success rate of 0.09%, and OpenAI's new Auto-review version from last week may change the picture. ##