How to enable Mythos 5 Claude Security scans on Enterprise

Claude Security is not a new chat model in the sidebar. It is a scan-scoped beta on Claude.ai: it reads a GitHub repository you own, surfaces vulnerabilities, and proposes patches for a human to review. Enterprise customers can run those scans on Claude Mythos 5. Interactive patching still uses the models your organization already has in Claude Code. Mythos does not become a general-purpose model, and it does not apply fixes on its own.

This walkthrough follows Anthropic’s official getting-started guide, help-center article, and product update so an Enterprise admin can flip the toggle, a premium-seat user can start a scan, and a reviewer can open a suggested patch in Claude Code on the web.

Confirm you are eligible

Before anyone clicks Scan, these requirements have to be in place:

An active Claude Enterprise account. Claude Security is in public beta for Enterprise only.
Claude Code on the Web enabled for the organization. Users reach it at claude.ai/code.
Extra Usage enabled, so consumption billing can run.
The Anthropic GitHub App installed on your GitHub organization, with access to the repositories you intend to scan. This is the same GitHub App used for Claude Code on the Web.
A premium seat for every user who will start scans. Standard seats do not include Claude Code on the Web.

If your team already uses Claude Code on the Web, most of this is already done. Confirm Extra Usage in Organization Billing, confirm the GitHub App under Installed GitHub Apps in GitHub, and confirm premium seats in the Claude.ai admin console.

Claude Security bills as token usage. The help center says scans are charged at direct token cost with no additional platform fee. The product update describes the same thing as standard token usage under your existing plan, with no separate add-on. Costs scale with the size and number of scans. After the feature is on, you can set a separate spend limit for it. Set those limits before the first large monorepo run.

Turn the feature on in the admin console

An organization owner enables Claude Security. In the official help article that path is Organization settings > Claude Security, then switch Turn on for your organization. The getting-started tutorial points at the same surface: claude.ai/admin-settings/claude-security.

Once that toggle is on and the prerequisites above are met, the Security Scan feature appears in Claude Code on the Web, and users see a Security entry point in the Claude.ai sidebar.

Start a Mythos 5 scan on a repo you own

Scans run from Claude Security, not from a free-form Mythos chat.

Go to claude.ai and click the Security icon in the left sidebar, or open claude.ai/security.
Choose a GitHub repository from the list of available repos. You can optionally scope the scan to a specific directory or branch. For larger repositories, Anthropic strongly recommends picking a directory so the scan is more likely to complete.
Choose scan effort if you are on a first scan or after material changes: Standard or Extended. Use Extended for a deeper review. Scans run on Claude Mythos 5.
Start the scan. Depending on repository size, it may take several minutes or hours.

You can run multiple scans at the same time and keep multiple working Projects. Separate Projects by repository, service, or team so findings stay attributable. Parallel scans help when you want to triage several repos at once or compare a hardened branch against main.

Only GitHub-hosted repositories can be scanned today. The published scope of use is explicit: scan code that you or your company owns and to which you hold the necessary rights. Do not point Claude Security at third-party or unrelated open-source repositories.

Read a finding before you touch a patch

When a scan finishes, findings show up with enough structure to triage. The product update lists a CWE (Common Weakness Enumeration) category, confidence and severity ratings, and a suggested fix. The getting-started guide also calls out vulnerability type, severity, affected file and line, and a description. The help center documents the finding record: title, details, location, impact, reproduction steps, recommended fix, severity (HIGH / MEDIUM / LOW), status (Open / Dismissed / Resolved), category, repository, branch, and date created. Dismissed findings also store a dismissal reason and an optional note.

Severity is assigned per finding based on exploitability in your codebase, not by category alone. The same class of bug can land at different severities in different repos:

High — exploitable by an unauthenticated remote attacker against a default deployment, with no meaningful preconditions.
Medium — exploitable behind authentication, or needs one or two realistic preconditions.
Low — needs three or more preconditions, local-only access, or lacks a concrete demonstrated attack path.

Severity is not configurable today. Start with the highest-severity items.

Example categories in the help article include injection (SQL, command, code, XSS, XXE, ReDoS), path and network issues (path traversal, SSRF, open redirect), auth and access failures (bypass, privilege escalation, IDOR/BOLA, CSRF, race), memory safety, cryptography mistakes, unsafe deserialization, and protocol or encoding confusion.

Scans are stochastic by design. Unlike a traditional static analyzer, Claude Security uses an agent that adapts its analysis each run. That is how it reaches for logic-level issues; it is also why two runs are not guaranteed to look identical.

You can export a Project, a Scan, or a Finding to CSV or Markdown. Dismiss false positives with a reason so they do not reappear and so later reviewers have an audit trail.

Open the suggested fix in Claude Code on the web

A finding is not a merged patch.

Click the remediation button on a finding to open a Claude Code on the Web session focused on that vulnerability. Claude analyzes the issue in context and generates a potential patch. Interactive patching uses whichever models your organization already has in Claude Code. The Mythos scan does not extend Mythos access to other surfaces.

Every patch must be reviewed and approved by a human before it can be implemented. Claude Security returns detailed findings rather than raw model access. Users receive those findings without talking to Mythos directly.

If you use the Claude Security plugin inside Claude Code, that plugin uses the models you already have there. Only the Claude Security product at claude.ai/security runs scans on Claude Mythos 5.

Keep scans operational

Once the first scan works, the official tutorial recommends scoping large repositories and monorepos to individual modules or subdirectories, scheduling scans on a cadence you will actually review (weekly is the example given), and naming an owner so the backlog does not become something people avoid. Export CSVs into the tracking system you already use and keep the files as an audit trail. You can also configure webhooks for scan-completion and new-finding events, and optionally set email notifications under Notifications in personal settings (General).

If claude.ai/security keeps asking you to install the GitHub App

That page runs a per-user check against your connected GitHub account. Organization-wide install can be fine while one person is still blocked. Check, in order:

A GitHub account is connected under Customize > Connectors.
That account is a member of the GitHub organization where the app is installed.
If the org requires SSO, the Claude app is authorized for it separately.
If the org uses an IP allow list, the per-user check is not covered by GitHub’s “Enable IP allow list configuration for installed GitHub Apps” setting. The help article says to add 160.79.104.0/21 as an organization-level allow-list entry.

What this toggle does not do

Enabling Claude Security does not give the team a Mythos 5 chat. It does not auto-merge patches. It does not scan non-GitHub hosts. It does not promise Zero Data Retention — Anthropic may retain data where required by law or to address Usage Policy violations. Treat it as a reviewer-backed scan of code you own, billed as tokens, scoped to Enterprise premium seats.

Open the getting-started tutorial next and walk the admin-console toggle, then run a directory-scoped scan on a repo you own and open one high-severity finding in Claude Code on the web. Do not apply the suggested patch until a human has reviewed it.