Claude Security Replaces Claude Code Security—but Its Hosted Beta Is Enterprise-Only

Claude Code Security has changed substantially since its February 2026 debut. The April 30 public-beta release renamed it Claude Security, made the hosted scanner available to Claude Enterprise customers and moved its scanning and patch-generation workflow to Claude Opus 4.7.

The broader access claim needs one important qualification. The current Claude Security product page lists the hosted service as an Enterprise public beta, while a separate Claude Code plugin is in beta for all Claude Code users and keeps code inside the user’s environment. The complete managed experience and the plugin are therefore related workflows, not identical versions with identical access conditions.

From restricted preview to an Enterprise security workflow

The original product was narrower in both name and availability. Anthropic’s February 20 launch record introduced Claude Code Security as a limited research preview for Enterprise and Team customers, offered expedited access to open-source maintainers, and stated that its team had used Opus 4.6 to find more than 500 vulnerabilities in production open-source codebases.

That vulnerability total describes Anthropic’s wider model-assisted security research, not the success rate of an ordinary customer scan. It does not establish how many findings were independently validated, how detection varied among repositories or how the system compared with specific commercial scanners under controlled conditions. Treating it as a universal benchmark would extend the claim beyond the evidence made public.

The later beta turned the initial repository review into a more repeatable operating process. Security teams can schedule scans, limit a scan to a directory or branch, preserve reasons for dismissing findings, export results in CSV or Markdown, and send results to Slack, Jira or other systems through webhooks. Those additions matter because vulnerability detection produces value only when findings can enter an organization’s existing triage, ownership and remediation workflow.

The underlying model also changed between releases. Opus 4.6 was associated with the research behind the February preview, while the Enterprise public beta introduced Opus 4.7 for finding vulnerabilities and preparing proposed fixes. The initial description remains relevant as launch history, but it no longer represents the product’s current name, model or availability.

What Claude Security actually does

Claude Security analyzes relationships across files and modules instead of relying exclusively on predefined signatures. It traces data movement, examines component interactions and looks for weaknesses whose significance depends on application context, including authorization failures, injection paths and business-logic errors.

Each candidate finding passes through a separate validation stage before it reaches a reviewer. The resulting entry can include the suspected vulnerability, its likely impact, reproduction information, severity and a confidence assessment. The system can also prepare a targeted patch that a developer opens in Claude Code for inspection and refinement.

Human approval remains part of the design. A proposed patch is not automatically equivalent to a safe fix, and the product documentation warns that Claude can make mistakes, particularly in critical systems. Confidence helps order an investigation; it does not prove that a finding is exploitable or that a generated change preserves all intended behavior.

Why this is not a replacement for conventional scanning

Claude Security and traditional static-analysis tools address overlapping but different problems. Rule-based scanners, linters, secret detectors and dependency checks are efficient at repeatedly finding known patterns, prohibited constructs and vulnerable package versions. Their predictable rules can also make results easier to reproduce between runs.

An agentic scanner is aimed at weaknesses that require broader interpretation of a particular codebase. It may connect behavior across several modules or identify a security consequence hidden inside otherwise valid operations. That flexibility also introduces variation: two scans need not explore precisely the same paths or produce identical findings.

The practical comparison is therefore not simply “AI versus SAST.” Claude Security can add a reasoning layer above established controls, while dependency analysis, secret detection, conventional static testing, threat modeling and regression tests continue to answer separate questions. Removing those controls would trade complementary coverage for dependence on one stochastic system.

Hosted service and plugin serve different access needs

The hosted Claude Security beta is designed for organizations on Claude Enterprise. Administrators enable access, users select an authorized repository or narrow the scope, and findings can flow into audit, notification and review systems. The managed workflow is intended to connect repository scanning, triage and patch review within the Claude environment.

The Claude Code plugin provides the broader beta route. It brings scanning, validation and patch preparation into Claude Code and is described as keeping code within the user’s environment. That distinction may be material for teams whose repository-handling rules do not permit the hosted workflow, although access to a beta does not by itself resolve an organization’s security, compliance or model-usage requirements.

Availability is not the only operational boundary. The hosted service currently requires Enterprise access, and suggested changes still need review before application. Teams evaluating either route must distinguish the convenience of generating a patch from the separate work of reproducing the vulnerability, testing the change and confirming that expected behavior remains intact.

What the public beta proves—and what it does not

The change since February is concrete: a restricted experiment became a named Enterprise product beta with scheduled scanning, targeted scope, documented triage and integrations, while the plugin opened a related in-environment workflow to the wider Claude Code user base. That is a meaningful expansion in product maturity and access, even though the hosted scanner remains Enterprise-centered.

The public material does not provide a universal detection rate, a controlled comparison against leading security products or a guarantee that generated patches are correct. It supports a narrower conclusion: Anthropic has productized repository-level AI security review and placed human-controlled remediation around it. Whether that produces better security outcomes will still depend on the repositories scanned, the quality of manual validation and the organization’s ability to act on credible findings.

Also read:

• Anthropic Explained When to Tinker with Effort Level in Claude Code (Spoiler: Not Often)
• A Data Room’s Security Depends on Configuration, Not Its Feature List
• Claude Tag: Anthropic Turns AI into a Persistent Teammate in Slack — and Andrej Karpathy Calls It the Third Major LLM UX Revolution