Anthropic has opened the beta of the Claude Security plug-in, which lets you use AI agents to scan QR codes and write update files.

Anthropic has turned Claude Security’s deep scanning capabilities into a Claude Code extension. Starting today, the beta is open, and all paid Claude Code users can install it. The extension runs directly within Claude Code conversations, with multiple Claude Agents dividing up the work to review the architecture, build threat models, and find vulnerabilities across files. Each finding and the corresponding patch code are also sent to a separate Agent to be re-checked. Users can scan the entire codebase, or only check branch differences, PR differences, or a single commit; after issues are found, they can choose to have it generate a patch, then review it themselves before deciding whether to apply it.
(Background: AI can use your passwords but can’t see them—Claude partnered with 1Password to build a zero-exposure access mechanism designed specifically for Agents)
(Additional context: How strong is AI at finding smart contract vulnerabilities? It earns 4.6 million per year, with costs as low as $1.22 per contract)

Key takeaways

• Anthropic has turned Claude Security deep scanning into a Claude Code extension. The beta opens today, and all paid Claude Code users can install it
• The extension runs within Claude Code sessions. Multiple Claude Agents review the architecture, build threat models, and find vulnerabilities; each finding and patch is re-checked by an independent Agent
• You can scan the entire codebase or only check branch/PR/single-commit differences. It can generate patches for manual review; scanning consumption counts toward Claude Code usage, and the hosted version is still limited to Enterprise

Programmers get an on-the-go security assistant. Anthropic has packaged the deep scanning capabilities of its Claude Security into a Claude Code extension. Starting today, it’s opening beta testing, and all paid Claude Code users can install it. This means that the automated vulnerability scanning that was previously tied to an enterprise product is now moved straight into the terminal workflow developers use every day.

It’s easy to use. Users can scan the entire codebase for a full health check, or do quick checks focused only on branch differences, PR differences, or a single commit. After it finds vulnerabilities, users can also choose to have it generate patches directly, then review them themselves and decide whether to apply them. In the end, whether to change anything is still up to humans.

Agents understand your code

The most critical part of this extension is that it runs within a Claude Code session. Behind the scenes, it’s not a single model running end to end; instead, multiple Claude Agents collaborate with different roles. They break down the program’s architecture, build threat models, and track issues across multiple files to uncover complex vulnerabilities that traditional pattern-matching tools often miss.

Even more important is the fail-safe mechanism: each finding and each patch is handed to an independent Agent for review again, preventing AI from generating false alarms or making changes arbitrarily on its own. According to Anthropic’s explanation, Claude Security targets high-severity vulnerabilities—such as memory corruption, injection attacks, authentication bypasses, and complex errors that require cross-file understanding of business logic to detect.

For the crypto community, this direction won’t feel unfamiliar. Automating the search for smart contract vulnerabilities with AI has long been a business. The difference this time is that Anthropic packages the same set of capabilities into a development workflow that developers are most familiar with, making security checks a convenient step alongside coding.

The extension is open to paid users; the hosted version is still only for Enterprise

Claude Security isn’t a brand-new product introduced only this time. It entered public beta for Enterprise customers as early as April 30, 2026, and before that it had already done a limited research preview under the name “Claude Code Security.” What’s new this time is the beta extension, lowering the threshold from Enterprise customers to all paid Claude Code users.

However, the separation still remains: the full hosted version of Claude Security is currently still only available to Enterprise customers, while general paid users receive the extension version. Also keep an eye on cost—the extension’s scanning consumption will be counted toward your Claude Code usage.

Common questions

How do I use the Claude Security extension? Who can install it?

The Claude Security extension is opening beta starting today, and all paid Claude Code users can install it. It runs within Claude Code conversations and can scan the entire codebase, or only check branch differences, PR differences, or a single commit. After it finds vulnerabilities, it can generate patches for human review and application, and scanning consumption is counted toward Claude Code usage.

How is Claude Security different from general vulnerability scanning tools?

It doesn’t rely on pattern-matching. Instead, multiple Claude Agents review the architecture, build threat models, and track issues across files to find complex vulnerabilities that traditional tools miss, and it also focuses on high-severity problems such as memory corruption, injection, and authentication bypass. Each finding and patch is also reviewed again by an independent Agent to reduce false positives.