Skip to content
Back/Claude Code

Claude Code Auto Mode

View in Graph
Updated 2026-10-24
3 min read
627 words

Claude Code Auto Mode

What it is

Claude Code Auto Mode 是 Anthropic 发布的四层权限自动化机制,在「每次手动确认」和「dangerously-skip-permissions」之间找到运行时语义判断的中间地带。四层模型包括:

  • allow:完全自主模式,agent 可自由执行任何操作
  • soft_deny:上下文敏感的 guardrail,根据请求合理性动态判断允许或拒绝
  • hard_deny:绝对禁止模式,特定操作永远不会被允许,不受上下文影响
  • environment:根据当前环境(repo 状态、分支保护规则等)动态调整权限

soft_deny 的典型行为:说 "force-push this branch" 会被拒绝,但补充 "because CI is failing and we need to unblock the team" 后会被允许。hard_deny 则始终拒绝,无论上下文多么合理。

Why it matters

Anthropic 内部数据显示,用户本就接受了 93% 的提示。手动确认大多时候只是疲劳操作,而非真正有效的安全闸。Auto mode 用低维护成本实现高自主性,把人类注意力从"机械点击批准"转移到"真正需要判断的异常"。

四层模型比二元"允许/拒绝"更接近真实安全需求:大多数操作应该被允许(allow),少数需要上下文判断(soft_deny),极少数永远禁止(hard_deny),环境规则提供兜底(environment)。

Key points

  • 权限设计的范式转移:从"列出允许做什么"转向"判断这次调用是否安全"
  • soft_deny 是核心创新:运行时语义判断替代静态规则,能捕捉"合法工具的恶意使用"(如删除远程 git 分支、上传 auth token)
  • hard_deny 是安全兜底:不受上下文影响的绝对禁止,防止模型被说服绕过关键限制
  • 两层防护架构:prompt-injection probe(防输入操纵)+ transcript classifier(防输出越界)
  • 与用户行为的收敛:93% 基线接受率说明自动化的边际风险远低于边际收益

Evidence across sources

Source Key Claim Relevance
AI Builders Digest 2026-05-05 模型分类器替代静态 allowlist;prompt-injection probe + transcript classifier;93% 基线接受率 官方技术实现与内部数据
AI Briefing 2026-05-17 Evening Daniel San Medium 文章详解四个 slots 的行为差异、配置技巧和 gotcha;soft_deny 的上下文判断逻辑与 hard_deny 的绝对禁止边界 用户视角:四层模型的实战配置指南

Open questions

  • Auto mode 的误报率(把安全操作拦下)和漏报率(放过危险操作)分别是多少?Anthropic 未公开
  • 当模型自身就是判断者时,是否存在"用模型攻击模型"的对抗路径?
  • soft_deny 的上下文判断窗口有多长?是否会因对话历史过长而产生漂移?
  • 这种运行时语义判断模式是否会成为所有 Agent 工具的默认权限设计?

Prompts for witness

  • 你使用 Claude Code 时,有多少比例的确认点击是"无脑通过"?哪些情况下你真的会停下来思考?
  • 如果 Auto mode 推广到所有 coding agent,你希望在哪些操作上保留强制人工确认?
  • soft_deny 的"根据上下文判断"是否让你更放心,还是更不安?为什么?
  • 在你的工作流中,哪些操作应该永远 hard_deny,哪些应该交给 soft_deny 判断?

2026-09-25:Auto Mode 默认化的团队 rollout 信号

来源:AI Briefing 2026-09-25 Evening

时间线中的团队 rollout 经验把 Auto Mode 的价值描述为减少个人规则、简化权限并降低 prompt injection 风险。这个信号补充了本页已有的运行时权限判断模型:默认化的关键不是减少所有保护,而是把配置变成团队级、可审计的边界。

Counterpoints & Gaps

  • 单条推文没有给出误报率、漏报率或实际事故数据。
  • 默认模式是否安全,取决于 hard deny、环境约束和审计链是否仍然有效。

Pending Jean review

  • claude-code/auto-mode — ai_draft; 在一个真实仓库上比较默认 Auto Mode、手动批准和严格拒绝的权限漂移与纠错成本。

2026-09-26:effort 与 prompt audit 成为权限之外的控制层

来源:AI Briefing 2026-09-26 Evening

本次时间线补充了两项与 Auto Mode 相邻的控制:prompt-audit 检查 CLAUDE.md、skills 和 agents 中已经不再需要的指令;effort 档位则按任务不确定性和验证负担调整模型投入。两者说明 agent 安全与效率不只由 allow 或 deny 决定,也取决于上下文是否精简、任务是否匹配适当的推理预算。

Counterpoints & Gaps

  • prompt-audit 不能替代权限审计,effort 也不能替代人工验收。
  • 社区经验没有给出不同 effort 档位在同任务集上的成本和质量差异。

Pending Jean review

  • claude-code/auto-mode — ai_draft; 对同一任务记录上下文清理、effort 档位、权限事件和最终接受结果。

2026-10-01 evening:effort、意图复述与执行可见性

来源:AI Briefing 2026-10-01 Evening

本次时间线补充了权限之外的三个控制点:用“Restate my intent before continuing”降低口述任务的意图漂移;用 coordinator/worker 的 effort 分层控制成本;用可见的执行状态和失败报告避免 agent 在不可见状态下消耗额度。

这些做法不能替代 allow、soft deny 或 hard deny,但说明安全与效率共同依赖任务边界、上下文可见性和人工验收。

Counterpoints & Gaps

  • effort 分层和意图复述的效果仍需在同一任务集上对照测量。
  • 单条产品体验反馈不能代表所有运行环境。

Sources (addendum)

Pending Jean review

  • claude-code/auto-mode — ai_draft;对同一任务记录意图复述、effort 档位、权限事件、成本和最终接受结果。

2026-10-19: Auto mode evaluation figures

Anthropic reports that Auto mode catches about 83% of “overeager” actions in its internal evaluation, against a 93% user-acceptance rate for approval prompts in the comparison workflow. These figures are specific to Anthropic's evaluation and should not be read as a general guarantee.

On Gray Swan's Agent Red Teaming benchmark, Anthropic reports prompt-injection success around 0.1% for a single attempt against Opus 4.7, rising to roughly 5–6% after 100 adaptive attempts. The gap illustrates why low single-attempt rates do not establish immunity under repeated adaptive probing.

Source: Anthropic, How we contain Claude; AI Briefing 2026-10-19 afternoon.

Pending Jean review: retain attribution and benchmark scope before using these numbers in operational policy.

2026-10-24: deterministic gates for high-impact actions

来源:AI Briefing 2026-10-24 afternoon

A user-described Auto Mode workflow kept script-level approval gates for money movement, posting, and account changes, even while enabling more autonomous operation elsewhere. The author found that a quote-repost tool had not been included in the protected-action list, added it, and shared the gate between Claude and Cursor. This is a concrete example of why an approval policy needs coverage across every tool path, not only a prompt or a single familiar action name.

Counterpoints & Gaps

  • This is one user's reported setup; the briefing does not independently verify the implementation or measure bypass coverage.
  • The example supports separating deterministic authorization from probabilistic classifiers, but it does not establish which actions every user should hard-block.

Sources (addendum)

Pending Jean review

  • claude-code/auto-mode — ai_draft; verify the reported tool-path coverage approach before treating it as a general policy.

Sources

Synthesized from 8 sources

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 8 sources.

    From AI Builders Digest 2026-05-05, AI Briefing 2026-05-17 Evening, 2026-09-25-22-58, 2026-09-26-22-57, 2026-10-01-22-59To Claude Code Auto Mode
    Sources: raw/briefing/AI Builders Digest/2026-05-05.md · raw/briefing/AI Briefing/2026-05-17-23-36.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-25-22-58.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-26-22-57.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-10-01-22-59.md · raw/briefing/AI Briefing/2026-09-25-22-58.md · raw/briefing/AI Briefing/2026-09-26-22-57.md · raw/briefing/AI Briefing/2026-10-01-22-59.md

Linked from