Felix Laurent
Anthropic released a new version of Claude Security on October 21, 2026. The scanner no longer waits for someone to start a scan. It runs one deep baseline pass, then keeps checking new changes and marks findings fixed when the patch ships. Anthropic says thousands of organizations have used the product since its public beta opened in April and found over 1M vulnerabilities with it. The update is in public beta for Claude Enterprise only.
What Anthropic announced
Claude Security started in February 2026 as a research preview called Claude Code Security. It opened as a public beta on April 30, running on Opus 4.7, with scheduled and directory-level scans.
The October version changes how the scanner keeps up with code. A team creates a project and connects its repositories. GitHub repos connect through the Claude GitHub app, GitHub Enterprise Server included. GitLab and other Git providers connect with an access token. Each repo gets its own branch and scope. Claude runs a baseline scan, then scans changes on a schedule Anthropic says is set by their risk and complexity.
One project can hold several repositories, so the scan can catch bugs that sit between services. Anthropic's example is a backend that trusts input the client was supposed to validate.
Fix tracking
Claude Security checks open findings against the code that changed. When the fix lands, the finding is marked fixed, whether the team patched it in Claude Code or somewhere else. If the same issue shows up again later, the finding reopens on its own.
Context and severity
Each project now has a knowledge base where a team describes how the code is deployed and which security policies apply. Claude Security also generates a threat model from the code, and the team can review and edit it. Later scans use that model to decide what to flag and what the environment already handles.
Every finding gets a severity rating judged against the project's threat model. It also comes with an explanation of how an attacker could reach the vulnerable code. Some findings come with a diagram of the attack path.
Integrations and access
The update also opens Claude Security to other tools and adds access control.
- An API to list and read projects and findings, resolve or reopen them, and create and manage projects and repositories. Personal access tokens see only what their user can
- Webhooks that fire when a new finding lands
- An MCP connector for Claude.ai and Claude Code: an engineer can pull a finding, start the fix and mark it resolved in one session
- Role-based access: admins set who can view projects, triage findings and create projects
The API and webhooks let a team feed findings into its own ticketing and alerting.
Models and cost
Anthropic says Claude Security runs on its full model family, including Claude Mythos, and does not require Cyber Verification Program access. Mythos-powered scans are available only in the Claude Security app on Claude.ai. A separate Claude Security plugin for Claude Code is in beta for all Claude Code users. It runs on the models in the user's Claude Code account, so it does not get Mythos.
Enterprise organizations get $25,000 in free usage credits a month for two months, on usage through December 31, 2026. Anthropic has not published a per-scan price.
What Anthropic did not publish
There is no detection-rate or false-positive figure and no independent test of the scanner. The 1M+ vulnerabilities number is Anthropic's own count, with no breakdown by severity. Customers quoted in the announcement include Vanta and IonQ. IonQ says the tool found material issues that a pen test run at the same time missed. Anthropic has not given a date for Team or Max access, which it said in April was "coming soon".
