Over a million vulnerabilities found by thousands of organizations since Claude Security entered public beta in April. Until now, every result was a snapshot: someone kicked off a scan, and the list started going stale with the next commit. As of October 21, 2026, Anthropic's scanner follows repository changes, knows which findings have been fixed, and prioritizes them against a threat model the team can edit.
It is in public beta for Claude Enterprise customers, with up to $25,000 a month in credits for two months.
What changes: from one-off scans to continuous coverage
The new workflow is built around projects. Each one groups up to eight repositories, and each repository is either in scope (scanned, with findings reported) or context-only (Claude reads it to understand the system but reports nothing in it). That matters more than it sounds: a backend that trusts validation the client was supposed to do only shows up when the scanner can see both sides.
After a baseline scan, which can take anywhere from minutes to hours, Claude watches new commits and decides what to scan based on the risk and complexity of each change. There is no fixed schedule.
The most useful part for whoever runs this day to day is the finding lifecycle. Claude Security checks every open finding against the code that changes and marks it fixed once the fix lands on the branch, whether it was written in Claude Code or anywhere else. If the issue comes back, the finding reopens on its own. That is the difference between a report someone has to reconcile by hand and a queue that maintains itself.
Severity is not a generic score either. Each project has a knowledge base describing how the code is deployed and which controls live outside it, and Claude generates a threat model the team reviews and corrects. Every finding comes with a severity rated against that model and the path an attacker would take to reach the vulnerable code, sometimes with a diagram. Vanta puts it this way: once they wrote their own risk policy into the threat model, findings matched how they prioritize "instead of only a severity score".
What it costs
There is no platform fee: you pay the direct token cost of scans, the threat model, project chat and fix checks. Claude Security can get its own spend limit, and each project shows its daily and month-to-date cost.
The launch promotion gives up to $25,000 a month in credits for two consecutive months. For organizations already on Enterprise, the two months start on October 21, 2026, and the offer ends on December 31, 2026 (UTC), so anyone signing up after November 1 won't get two full months. Credits are drawn before committed spend and don't roll over.
Getting it running
- An organization owner turns it on under Organization settings > Claude Security. Each user needs a premium seat, and the organization needs usage credits enabled.
- Connect repositories: on GitHub.com or GitHub Enterprise Server, by installing the Claude GitHub App; on GitLab, Bitbucket or other Git servers, with the repository URL and a token.
- For large monorepos, scope the project to a folder and exclude the ones you don't care about.
- Fill in the Knowledge tab with how the service is deployed, and review the generated threat model before trusting the severities.
- Hand out Full, Triage or View roles, and use RBAC to decide who can create projects.
There are three ways to plug it into existing tooling: an API to list and read projects and findings, resolve or reopen them and manage repositories, authenticated with a personal access token that can only see and change what its owner can; webhooks that fire on every new finding (to Slack or Jira, for instance); and an MCP connector for Claude.ai and Claude Code, so you can open a finding, ask for the fix and mark it resolved in the same session.
Gotchas and edge cases
- Two scans of the same code can return different results, as the support docs themselves warn. Treat it as an extra layer, not as a deterministic CI gate that blocks merges.
- No Zero Data Retention. For teams whose contracts require it, this is the deciding factor.
- Inbound traffic: a self-hosted Git server has to accept connections from Anthropic's IP ranges.
- Migrating from the previous version does not copy members, webhooks, excluded folders, scan history or scheduled scans, and GitHub Enterprise projects can't be migrated. Credits only cover projects in the new experience.
- Eight repositories per project at most: with lots of microservices, you have to choose which ones belong together.
If I already had a project in the old version, I would start a new one with the most internet-exposed service and its dependencies as context, and get the threat model properly written before the spend stops being covered by credits.
Sources
- Claude Security now scans continuously and tracks fixes, Anthropic, October 21, 2026.
- Use Claude Security, Claude Help Center.