A Look at Vercel's Security Tool Deepsec

Vercel has open-sourced deepsec. It's a security harness that lets coding agents explore a codebase, running on your own infrastructure so sensitive source code doesn't have to be handed over to an external cloud. Inference is handled through your existing Claude or Codex subscription, and for large repositories work can be fanned out to Vercel Sandboxes for parallel execution. Inside Vercel, scans have used as many as 1,000+ concurrent sandboxes.

The workflow has five stages.

• Scan: Sweeps all files with regex to pick out security-sensitive files.
• Investigate: An agent traces data flow and mitigations for candidate files.
• Revalidate: A second agent verifies the results to reduce false positives and re-score severity.
• Enrich: Uses Git metadata and optional services to identify who's responsible for the fix.
• Export: Exports results in a directive format that's easy to move into human-facing tickets.

It also supports Anthropic's and OpenAI's cyber versions of their models, but the default models work fine too. In internal testing, a classifier was inserted after each stage to check for refusals, and with Opus 4.7 and GPT 5.5, refusals were effectively a non-issue.

Vercel acknowledges a false-positive rate of roughly 10-20%, which is why the revalidate stage exists. The plugin system lets you attach custom scanners per codebase, and regex matchers tailored to auth paths, data layers, and team-specific rules are key. Vercel has validated this on its own monorepo and on open-source code from customers and partners, and on the Dub codebase it even built a custom scanner plugin covering every auth path.

Getting started is a single line: npx deepsec init. This creates a ./.deepsec directory at the repository root, under which configuration and investigation records accumulate. Vercel notes that this is still early-stage development and is taking feedback and contributions on GitHub.