Anthropic's New Claude Security Plugin Uses 3-Way AI Panel to Hunt Bugs

Anthropic just shipped a security plugin for Claude Code that’s worth paying attention to — not because it finds bugs, but because of how it decides what counts as a real finding. The beta release, available now from the official plugin marketplace, drops a single /claude-security command into your terminal session and sets off a multi-agent vulnerability scan that reads more like a formal audit than a typical SAST tool’s spray of false positives.

The scan pipeline is a six-phase orchestration script written in JavaScript. It starts by inventorying your repository into components, builds threat models for each one, then fans out researchers across four fixed categories: injection, auth, memory safety, and crypto. Here’s where it gets interesting. Every candidate finding must survive a three-lens adversarial panel — reachability, impact, and defenses — with each lens returning a structured TRUE_POSITIVE or FALSE_POSITIVE verdict. You need at least two of three votes to keep a finding. If the panel is unanimous, the finding gets a high confidence ceiling; a 2/3 split caps it at medium. The tally is computed in Python by the report renderer, not asserted by the model that found the issue, so you can actually verify the report’s account of its own rigor.

Anthropic didn’t skimp on the operational details either. You choose from four effort tiers — low, medium, high, or max — which control how many components get scanned (capped at 12 for low and medium, 24 for high and max), how many researchers work each cell, and how many gap-fill sweeps run. For small diffs or limited scopes, the whole thing condenses into a single-researcher config. The orchestrator runs on Opus, while the cartographer and code explorer use Sonnet, and all scan agents are locked to read-only tools. That’s the kind of architectural transparency that makes security engineers actually trust a tool.

What you get on disk is refreshingly straightforward: a timestamped directory with a human-readable markdown report, a JSONL file with the same findings, and a revision stamp that tells you exactly which commit was scanned at what effort level. Each finding includes the CWE ID, the exact sink line, exploit scenario, preconditions, and a recommendation. You can then turn selected findings into .patch files that you review and apply yourself. It’s not a magic fix button — it’s a structured argument about what’s vulnerable and why, with the receipts to back it up.