Anthropic Releases Claude Security Plugin Beta, Supporting Terminal Code Change Scanning and Patch Generation

On July 22, 2026, Anthropic officially announced the beta release of the Claude Security plugin. Integrated within Claude Code, this plugin allows users to scan code changes for vulnerabilities directly from the terminal, or perform multi-layered deep scans across the entire codebase, offering vulnerability explanations and remediation suggestions. Users can conduct security reviews using their existing Claude subscription without additional tools.

Scanning Workflow and Technical Implementation

The plugin opens a menu via the single command /claude-security, offering three task options: scanning the entire codebase or a specified subset, scanning differences in branches or pull requests, and converting findings into .patch files. Installation requires two commands, with the plugin source code publicly available in the claude-plugins-official repository, current version 0.10.0.

Scanning employs a dynamic workflow, where a JavaScript orchestration script assigns sub-agents to execute six phases: component partitioning, threat modeling, taxonomy research, gap filling, three-perspective adversarial validation, and best-effort review. Research targets four fixed categories: injection and input, authentication and access, memory and unsafe operations, and encryption and keys. Components written in memory-safe languages can omit the memory category. Run sizes range from low, medium, high, to maximum, with the number of components, researchers, and sweep iterations increasing accordingly. Small-scale scans simplify to a single-researcher mode.

Integration with Existing Development Workflows

The plugin runs multi-agent vulnerability scans within Claude Code sessions, converting selected findings into patch files for user review and application. It can perform comprehensive scans of full repositories or check terminal changes before commits. All operations use the developer's existing Claude inference resources, eliminating the need to switch to a separate application or submit code to an external system.

The plugin supports scheduled or targeted scans, allows findings to be dismissed with reasons for audit trails, and provides CSV and Markdown export. Webhooks can directly integrate with enterprise tools such as Slack and Jira. Technology partners include CrowdStrike, Microsoft Security, Palo Alto Networks, among others, while service partners include Accenture, BCG, Deloitte, and more.

Practical Impact on Stakeholders

For developers, security checks are embedded within existing terminals and Claude Code sessions, reducing context switching. When reviewing pull requests or verifying generated code, they receive contextual explanations and targeted remediation instructions directly. Developers must review all suggested changes themselves, especially in sensitive scenarios like authentication, encryption, and database queries.

For enterprise users, Claude Enterprise customers can access the plugin via the Claude.ai sidebar or claude.ai/security, requiring no API integration or custom agent construction. Organizations can use AI scans as a supplement to, rather than a replacement for, existing static analysis, dynamic testing, and penetration testing. The material does not provide details on language coverage, accuracy rates, or CVE mapping.

For security teams, early detection can reduce the likelihood of issues being introduced later in the CI/CD pipeline. Conventional static tools often generate a high number of false positives.

Strategic Observations

Based on publicly available facts, Claude Security ties security scanning directly to Claude subscriptions and the Claude Code workflow, lowering the adoption barrier. During the beta phase, user feedback on false positive rates, patch applicability, and actual integration performance with existing security platform webhooks will determine whether it truly embeds shift-left practices, rather than merely providing an additional scanning channel.

If future versions expand to Claude Team and Max customers, coverage will further increase. For enterprise selection, it is recommended to pilot on non-production repositories, recording scan duration, number of findings, and manual review effort, before deciding whether to incorporate it into formal development workflows.