Claude Code managed settings: verify your team's fix

The short answer

Claude Code's managed-settings advisory concerns organization policy delivery when a stored API key interferes with Team or Enterprise authentication. S1 Update affected clients to 2.1.260 or later, then check that the intended policy reaches each device. S1S3 Treat the software update and the policy check as separate acceptance steps: record the running version first, then compare the effective permissions with an owner-approved policy reference.

Am I affected?

Use the vendor's plan-specific starting versions, rather than applying the broad package range indiscriminately. S1

For a safe starting inventory, run claude --version in the environment where developers actually use the tool; it reports the installed version. S4 Ask the organization owner to identify the intended policy delivery method. Avoid opening credential stores merely to complete this checklist.

Create one row per device or work environment, with these fields: responsible person, account plan, reported version, delivery method, verification date and unresolved questions. Keep credentials, project contents and full diagnostic logs out of that worksheet. If the delivery method is unknown, assign someone to establish it before marking the row complete.

Do not interpret an unaffected policy-delivery method as a general security assessment of the device. Keep the conclusion tied to this advisory's scope.

What happened

The vendor describes a credential-selection error that could leave server-delivered policy absent or stale, despite an organization-authenticated session. S1 The issue is identified as CVE-2026-103012 and GHSA-gfvf-j8jh-jxxw. S1

Managed policy is intended to govern capabilities such as tool access and file permissions across an organization's Claude Code users. S2 That makes policy delivery worth checking as a rollout requirement, alongside the ordinary version inventory.

Use the timeline to distinguish disclosure from verification. The second row records this guide's limited metadata check; it is not a claim about when the fix was released, whether an organization received it, or whether anyone exploited the issue.

How to fix it

Developer-managed installations

• Record the current version and the installation method before changing anything.
• For installations managed by Claude Code's own updater, use claude update, then restart and check the version again. S4
• For Homebrew, use the upgrade command for the installed cask: brew upgrade claude-code or brew upgrade claude-code@latest; for WinGet, use winget upgrade Anthropic.ClaudeCode. S4
• Hand the version result to the person completing the policy-delivery check below. Keep the task open until that person records a result.

Prefer the existing installation method. Do not add a second installation simply to make a version command print a different number. If the result is unexpected, ask the device administrator to resolve which installation the developer is using before continuing.

Centrally managed teams

Assign one rollout owner and one policy reviewer. Prepare a small pilot group that includes the installation methods your team actually uses. Ask each participant to supply the worksheet fields above, then review the results before expanding the rollout.

Define completion in writing: the intended client version is running, the expected delivery method is identified, and the current permission list has been compared with the owner's reference. Record exceptions individually. Do not turn an unresolved device into a team-wide success by averaging it into a rollout percentage.

If you cannot update today, do not invent a credential-deletion workaround. The advisory recommends updating manual installations and says standard auto-update users have already received the correction. S1 Escalate blocked devices to the administrator responsible for software delivery.

Verify the fix

Run the acceptance check with an authorized user in their normal work environment. Follow the vendor's documented sequence: restart Claude Code, inspect the remote-settings result in claude doctor, and use /permissions to review effective rules. S3

Use this decision worksheet to turn those observations into a clear next action:

The vendor documentation describes separate outcomes for loaded settings, absent organization configuration, failed retrieval and skipped retrieval. S3 A useful verification record should preserve that distinction rather than just say “doctor passed.”

For example, use a ticket with two independently completed fields: client update and policy delivery. Name the person who checked each field and record the time. This is a suggested review process, not a report of testing on a customer device.

We ran npm view @anthropic-ai/claude-code@2.1.260 version engines.node against a fresh, locally cached registry response in offline mode. It returned version 2.1.260 and a Node requirement of >=22.0.0. S5 This corroborates package metadata only; it does not verify an installation, authenticate an organization or test policy enforcement.

What we know and what we don't

The advisory requires local access and a stored key; an empty policy cache is an additional condition for its no-policy outcome. S1 It does not establish that a particular developer, team or customer was compromised. S1

Keep four questions separate in your incident notes: whether the version falls within the range, whether the relevant configuration existed, whether policy delivery failed, and whether an unauthorized action occurred. Assign unknowns explicitly. Do not use the version result alone to answer the other questions.

This guide's original contribution is the plan-specific applicability table and the two-part acceptance worksheet. We have not reproduced the defect or inspected customer accounts. Treat the guide as a way to organize authorized local verification, with the advisory as the source of vulnerability facts.

Questions developers are asking

An older developer report describes successful organization authentication alongside missing managed rules, and says signing out and back in did not restore delivery. S6 Those observations motivate the questions below; the report is not evidence that this newly disclosed issue caused that user's problem.

Does signing in to the right organization prove policy loaded?

Record sign-in and policy delivery separately. The reported experience shows why developers ask this question, but it does not diagnose every failed sync. S6 Ask the owner to compare the current effective rules with the intended reference before completing the policy field in your ticket.

Should I keep signing out and back in?

Do not use repeated sign-in attempts as your acceptance criterion. The developer report says that did not resolve its missing-policy symptom. S6 Complete the version update and documented delivery check, then give an unresolved case to the administrator with a concise description of the observed state.

Where should I compare the rules my team expects?

Use /permissions for the effective permission list, as the vendor's delivery-verification instructions recommend. S3 Have the policy owner identify the reference to compare against. Record differences and ownership of the follow-up; avoid copying sensitive paths or entire configuration files into a public issue.

Updates

• 2026-09-30 07:28 UTC — We prepared this initial guide with vendor scope, plan-specific ranges and a recorded package-metadata check. No publication or customer-device verification is claimed.