Software development has shifted rapidly toward AI-assisted tools that generate, edit, and review code at high speed.
This change has improved productivity for many teams while also raising the volume of code that must be examined for security weaknesses before it reaches production. Traditional scanners continue to play a role, yet they often struggle with context-dependent issues that cross file boundaries or involve subtle logic errors.
Anthropic released the 'Claude Security plugin' for Claude Code in public beta on July 22, 2026.
The tool allows developers to scan recent code changes or an entire repository for high-severity vulnerabilities directly from the terminal, using the Claude inference already in use. Scans can run before a commit or across a full codebase without requiring a separate external service beyond the existing model access.
The plugin operates through a multi-agent process inside a Claude Code session. Agents map the code architecture, construct threat models, search for vulnerabilities, and then independently verify findings to limit false positives.
It examines Git history, traces data flows across files, and considers business logic rather than relying solely on pattern matching.
This approach targets issues such as memory corruption, injection flaws, authentication bypasses, and complex logic errors that often span multiple files and are frequently missed by conventional rule-based scanners.
Findings include severity classifications and confidence rankings. Suggested patches are generated to match the project's coding style but require manual human review and are never applied automatically.
Read: Claude Security Enters Public Beta With AI-Powered Code Vulnerability Scanning And Patch Suggestions
In particular, Claude Security focuses on serious vulnerabilities that most traditional tools usual miss. They include:
- Memory corruption.
- Injection flaws.
- Authentication bypasses.
- Complex logic errors.
Installation uses the command /plugin install claude-security@claude-plugins-official followed by /reload-plugins, provided Claude Code version 2.1.154 or later is running on a paid plan with the necessary features enabled.
The main interface is accessed via /claude-security, which offers options for full scans, change-based scans (such as branches, commits, or diffs), or patch generation. Results are written to a timestamped directory containing Markdown reports and JSONL data.
For larger repositories, scoping the scan to specific directories or layers is practical to manage token usage and runtime. The plugin can integrate findings into existing workflows through webhooks for tools like Slack or Jira, or via CSV and Markdown exports. Documented dismissals persist across subsequent scans.
This release sits alongside earlier Anthropic offerings, including a lighter security-guidance plugin that performs real-time pattern checks and reviews during editing and commits, and a web-based Claude Security interface available to enterprise customers.
The terminal plugin brings similar reasoning capabilities into the daily coding environment for a broader set of Claude Code users. It is designed to complement existing static analysis tools rather than replace them.
The feature remains in beta.
Full repository scans can increase token consumption and may require more capable models, making scoped or incremental scans preferable in many cases.
Results are nondeterministic, so repeated scans of the same code can produce varying findings. False positives and missed issues remain possible despite the verification step. Suggested patches should be treated as drafts that need careful review, particularly for critical systems.
In other words, the tool does not substitute for established application security practices, penetration testing, or dedicated analysis pipelines.
However, by placing researcher-style analysis inside the terminal environment where code is written, the plugin shortens the interval between introduction of a potential issue and its detection.
Open-source maintainers and teams without dedicated security resources gain an additional review layer with relatively low setup friction, while organizations can use it as one more checkpoint within larger processes.
