For years, security testing has been something that happens after the code is written — sometimes long after. Developers push changes, SAST tools scan for known patterns, and security teams work through backlogs that never seem to shrink. It’s a reactive model, and attackers have gotten very good at exploiting the gaps it leaves open.
Anthropic is trying to change that dynamic with Claude Code Security, a new capability built into Claude Code on the web. Launched in February 2026 as a limited research preview, it provides development and security teams with an AI-powered way to find and fix vulnerabilities — including subtle, logic-level ones that traditional tools routinely miss.
More Than Pattern Matching
Most static analysis tools work by matching code against known vulnerability signatures. That approach catches the obvious stuff: Exposed credentials, outdated encryption, basic injection flaws. But it struggles with anything more nuanced, such as broken access controls, flawed business logic, or multi-component vulnerabilities that only become apparent when you trace data across an entire application
Claude Code Security takes a different approach. Rather than matching patterns, it reads and reasons about code the way a security researcher would — understanding how components interact, following data as it moves through the application, and surfacing issues that require context to spot
Anthropic’s own Frontier Red Team tested this capability against production open-source codebases. Using Claude Opus 4.6, the team found more than 500 vulnerabilities that had gone undetected — in some cases for decades — despite years of expert review. They’re currently working through responsible disclosure with the affected maintainers.
Built to Reduce Noise
Security tooling has an alert fatigue problem. When tools flag too many false positives, analysts start tuning things out — and real issues get buried in the noise.
Claude Code Security addresses this with a multi-stage verification process. Before a finding reaches an analyst, Claude re-examines the result, attempting to prove or disprove its own conclusion. Findings that pass verification are assigned severity ratings and confidence scores, enabling teams to prioritize what matters most.
Validated findings surface in the Claude Code Security dashboard, where teams can review the finding, inspect a suggested patch, and approve the fix. Nothing is applied automatically. Developers and security engineers stay in the loop at every step.
“Detecting vulnerabilities as code is written moves security to the point of origin, where reasoning about how components interact catches logic-level flaws that pattern-matching scanners push to a later stage,” said Mitch Ashley, VP and practice lead for software lifecycle engineering and AI-native software engineering at The Futurum Group. “If you are not solving security at the point of origin, you continue to create a backlog downstream for other tools and people to clear. As AI accelerates how much code ships, that deferred work compounds faster than teams can absorb.”
The Supply Chain Gap
One thing worth noting: Claude Code Security focuses on source code analysis. That’s a meaningful capability, but it doesn’t cover the full attack surface of modern software. Supply chain threats — malicious packages, compromised dependencies, tampered build artifacts — require a different set of tools. Teams should treat Claude Code Security as a single layer within a broader security strategy, not as a replacement for supply chain monitoring or runtime protection.
That said, for the class of vulnerabilities it does target, the reasoning-based approach has real advantages over rule-based scanners. Findings that require understanding how a middleware component interacts with an authentication layer, for example, are exactly the kind of logic-level issues where pattern matching falls short.
Who Can Access It
The limited research preview is currently open to Enterprise and Team customers. Anthropic is also offering free, expedited access to open-source maintainers who apply directly. The goal is to collaborate with early users to refine the tool’s capabilities before broader rollout.
Because it’s built on top of Claude Code, security teams can review findings and iterate on fixes inside the same environment where development is already happening. That’s a meaningful workflow improvement — fewer context switches, less friction between finding a bug and shipping a fix.
Timing Matters
This capability arrives at an inflection point. AI-assisted coding is accelerating the pace of software development, and that means more code, more often, with potentially more opportunities for vulnerabilities to slip through. Research from Veracode found that 45% of AI-generated code samples failed security tests and introduced OWASP Top 10 vulnerabilities during evaluation.
The same AI capabilities that can introduce vulnerabilities can also find them. Claude Code Security is Anthropic’s attempt to put those capabilities in the hands of defenders — not just at the end of the development cycle, but throughout it.
Security has always been easier to address early. The question is whether teams will use tools like this to shift that work left, or keep treating it as someone else’s problem until it isn’t.