Anthropic Launches Claude Security Beta: AI-Powered Vulnerability Scanning in Terminal

Anthropic has pushed the Claude Security plugin for Claude Code into public beta, letting developers scan for vulnerabilities before every commit without ever leaving their terminal.

The move extends a capability that debuted as a limited research preview back in February, when Anthropic pointed Claude Opus at production open-source codebases and surfaced more than 500 high-severity vulnerabilities that had survived decades of expert review.

The plugin builds on Claude Code Security, which Anthropic first shipped as a research preview to Enterprise and Team customers in late 2025, with expedited access for open-source maintainers.

Anthropic Launches Claude Security Beta

That preview evolved into a full public beta on April 30, 2026, initially reserved for Enterprise customers before wider Team and Max access followed.

The newest iteration folds the same scanning engine directly into Claude Code’s command line, running on the Claude inference teams already pay for rather than requiring a separate API integration or custom agent build.

Traditional static analysis tools flag known signatures; Claude Security instead traces data flows across files and generates hypotheses about how untrusted input moves through an application, the way a human security researcher would.

Anthropic says this approach is what lets it catch context-dependent flaws, memory corruption, injection vulnerabilities, authentication bypasses, and complex business-logic errors that pattern-matching scanners typically miss.

The plugin’s pipeline is built around three distinct phases, and the middle step is arguably its most consequential design choice for reducing analyst fatigue.

Scan: Claude runs a multi-agent pass across a repository, a diff, or a targeted directory, tracing multi-component vulnerability patterns rather than checking files in isolation.
Validate: every finding goes through an adversarial verification pass where Claude challenges its own conclusions, a mechanism Anthropic credits with cutting false positives and surfacing more genuine issues.
Patch: confirmed findings ship with a targeted, style-consistent fix that a human must review and approve before it merges, keeping remediation authority with the security team.

Rather than replacing existing tooling, the plugin is designed to slot into workflows security teams already run. Scans can be scoped to a single directory or branch, scheduled to run recurringly for ongoing coverage, and findings can be exported to CSV or Markdown or pushed via webhook into Slack and Jira for triage.

Dismissed findings are also documented so triage decisions carry forward across future scans, addressing a common complaint about noisy scanners that force teams to re-litigate the same false positives repeatedly.

Claude Security’s public beta launched for Enterprise customers on April 30, 2026, with Team and Max plan access following in subsequent weeks. It runs on Anthropic’s most capable generally available models, the same ones the company says it uses to secure its own codebase, and requires no custom infrastructure to deploy.

The GitHub Action variant, which posts inline PR comments with security findings, has been available separately since mid-2025 for teams wanting review gates earlier in the CI/CD pipeline.

For a threat intelligence audience already tracking AI-assisted vulnerability discovery, this beta signals Anthropic’s clearest bet yet that reasoning-based scanning, not just faster pattern-matching, will define the next generation of application security tooling.