LIVE
Anthropic says Claude helped trace the origin of NIST's mysterious ECDSA seeds16/10/26 · Anthropic|Reflection AI unveils Beam, a 501-billion-parameter open model16/10/26 · Reflection AI|Meta releases Muse Spark Open, a mixture-of-experts model built for local Mac Studio inference15/10/26 · Meta|Meta releases Muse Spark Open, an open-weight multimodal model15/10/26 · Meta|Google DeepMind outlines safety measures behind Gemini 4 Argon13/10/26 · Google DeepMind|Tev-Gym: A Training Recipe to Fix Sparse Rewards in Long-Horizon Agents13/10/26|Google DeepMind references Gemini 4, code-named Argon13/10/26 · Google DeepMind|Google DeepMind unveils Gemini 4 Argon, a robotics-focused variant13/10/26 · Google DeepMind|Character.AI files for Chapter 11 bankruptcy, sells assets to Disney13/10/26 · Character.AI|Study Shows AI Models Can Provably Conceal Arbitrary Capabilities13/10/26|NVIDIA RTX Spark laptops bring local AI to premium Windows PCs17/10/26 · NVIDIA|OpenAI discloses four more incidents involving sensitive data17/10/26 · OpenAI|Anthropic says Claude helped trace the origin of NIST's mysterious ECDSA seeds16/10/26 · Anthropic|Reflection AI unveils Beam, a 501-billion-parameter open model16/10/26 · Reflection AI|Meta releases Muse Spark Open, a mixture-of-experts model built for local Mac Studio inference15/10/26 · Meta|Meta releases Muse Spark Open, an open-weight multimodal model15/10/26 · Meta|Google DeepMind outlines safety measures behind Gemini 4 Argon13/10/26 · Google DeepMind|Tev-Gym: A Training Recipe to Fix Sparse Rewards in Long-Horizon Agents13/10/26|Google DeepMind references Gemini 4, code-named Argon13/10/26 · Google DeepMind|Google DeepMind unveils Gemini 4 Argon, a robotics-focused variant13/10/26 · Google DeepMind|Character.AI files for Chapter 11 bankruptcy, sells assets to Disney13/10/26 · Character.AI|Study Shows AI Models Can Provably Conceal Arbitrary Capabilities13/10/26|NVIDIA RTX Spark laptops bring local AI to premium Windows PCs17/10/26 · NVIDIA|OpenAI discloses four more incidents involving sensitive data17/10/26 · OpenAI|
+ nAIvigateA SHEET FOR PEOPLE WITHOUT 2 HOURS
+ nAIvigate · fiches / claude-code-entreprise-modes-de-permission
▣ fiche-007.md ×
□ index.md
☰ ContentsVerdictThe six modesWhich mode, for whomFive enterprise settingsThe classifier and its limitsAction plan

Claude Code now starts in auto mode. Who decided that in your company?

Claude Code reads, writes and runs commands on the machine. The permission mode decides what it does without asking. Since version 2.1.283, the starting mode is auto: a second model reviews actions in your place. The documentation is explicit: it reduces prompts, it does not guarantee safety. Across a team, that choice should not be left to a default.

Verdict

the essentials in 10 seconds4 rows
TEST ↗Is your configuration safe? Paste your settings file into the checker: a score and the corrected file, with nothing sent.FREE TOOL
BY DEFAULTSince version 2.1.283, terminal and VS Code sessions start in auto mode, on every plan and every provider.CHANGE
CLASSIFIERA second model reviews actions before they run. It reduces prompts, it does not guarantee safety.LIMIT
ENTERPRISEThe starting mode, auto mode and the no-check mode are set in managed settings. Someone still has to do it.TO CONFIGURE

The six modes

what runs without asking6 rows
manualReads only without a prompt, everything else is asked. Its config value is written default.SENSITIVE
acceptEditsReads, file edits and common commands (mkdir, mv, cp) inside the working directory.
planClaude explores and proposes a plan. No changes until you approve it.EXPLORE
autoEverything runs, with background safety checks. It is the starting mode since 2.1.283.DEFAULT
dontAskReads and pre-approved tools only. Anything that would prompt is denied.CI
bypassPermissionsEverything, with no checks. For isolated containers and virtual machines only.ISOLATED ONLY

Which mode, for whom

a decision, not an inherited setting5 rows
sensitive dataA machine that touches production, secrets or customer data: manual, and plan to explore before changing anything.
developmentCode reviewed by a human afterwards: acceptEdits, or auto once trusted infrastructure is declared.
long tasksauto, with deny rules on what must never happen and a sandbox as reinforcement.
continuous integrationdontAsk with an explicit list of allowed tools. The session never waits for an answer.
unattendedbypassPermissions only inside a container or virtual machine, never as root, never on a workstation.CONTAINER

Five enterprise settings

to put in managed settings5 rows
defaultModepermissions.defaultMode sets the starting mode of terminal sessions across the organisation. People can still switch to auto.
disableAutoModepermissions.disableAutoMode set to disable removes auto mode: nobody can select it.KILL SWITCH
disableBypasspermissions.disableBypassPermissionsMode set to disable blocks the no-check mode.
denydeny rules block in every mode, bypassPermissions included. They are the only hard guarantee.GUARANTEE
environmentautoMode.environment declares your trusted repositories, storage and services. Without it, the classifier treats everything else as external.

The classifier and its limits

what it does, what it does not5 rows
blocksDownloading then executing code, sending sensitive data outside, deploying to production, force pushing, destroying files that existed before the session.BY DEFAULT
allowsLocal operations in the working directory, installing declared dependencies, read-only requests, pushing to the repository you work in.
instructionsA boundary stated in the conversation ("don't push") is honoured, but it can be lost when the context is compacted. For a guarantee, use a deny rule.TRAP
costOn Enterprise plans and API accounts, classifier calls count toward token usage.
fallbackAfter 3 blocks in a row or 20 in total, auto mode pauses and prompts come back.

Action plan

to start this week6 rows
01Check installed versions: from 2.1.283, the starting mode changed without any action on your side.FIRST
02Choose the starting mode per population and set it in managed settings.
03Write the deny rules: production, secrets, deletions, publishing.
04Block bypassPermissions outside containers.
05Declare trusted infrastructure if auto mode is kept.
06Train the teams: what each mode does, and why a spoken instruction is not enough.

Go deeper

fundamentals, in Learn3 links
Is your configuration safe? Paste your settings file: a score, the list of gaps and the corrected file, with nothing sent. Then, training for your teams.
TEST MY CONFIGURATION →
⎇ mainfiche #007 · securitesources: Documentation Claude Code : Choose a permission mode (consultée le 05.10.2026)
Follow nAIvigate | naivigate.com/en/fiches