# deepsec

> An open-source, agent-powered vulnerability scanner that runs on your own infrastructure to find hard-to-detect security issues in large codebases.

deepsec is an open-source security harness built by Vercel Labs that uses coding agents to perform deep vulnerability scanning on large codebases. Released in May 2026 by Vercel CTO Malte Ubl, it runs entirely on your own infrastructure — no cloud service required for source code access — and leverages your existing Claude or Codex subscription for inference.

## What It Is

deepsec is an agent-powered vulnerability scanner designed to surface hard-to-find security issues that have been lurking in applications for a long time. Unlike traditional static analysis tools, it combines regex-based candidate identification with AI-driven investigation using frontier models at maximum reasoning levels (Claude Opus 4.7 and GPT 5.5). The tool is written in TypeScript, licensed under Apache 2.0, and distributed via npm.

## How the Pipeline Works

The scan workflow proceeds through distinct stages:

- **Scan**: A fast regex-only pass over all files to identify security-sensitive areas as candidates for deeper investigation.
- **Investigate**: Coding agents examine each candidate file, tracing data flows, checking for mitigations, and producing findings with severity ratings.
- **Revalidate**: A second agent pass validates findings to remove false positives and reclassify severity. The blog post notes a false positive rate of roughly 10–20% before revalidation.
- **Enrich**: An agent uses git metadata and optional plugins to identify contributors responsible for fixing each issue.
- **Export**: Findings are formatted as actionable instructions that can be turned into tickets for humans or coding agents.

A `process --diff` mode supports PR review and CI gating by scanning only files changed in a diff.

## Distributed Execution via Vercel Sandboxes

For large monorepos, deepsec supports optional fanout to Vercel Sandbox microVMs for parallel remote execution. The blog post states that scans on Vercel's own codebases routinely scale up to 1,000+ concurrent sandboxes. The local working tree is tarballed and uploaded; `.git` is excluded. Both OIDC tokens (local) and access tokens (CI) are supported. If a run is interrupted, deepsec picks up where it left off, skipping already-analyzed files.

## Plugin System and Customization

deepsec ships with a plugin system for adapting scans to specific codebases. The most common plugins are custom scanners — regex matchers tuned to an application's auth model, data layer, or team conventions. The documentation recommends using a coding agent to write custom matchers based on findings from an initial scan. The tool works best for applications and services; libraries and frameworks may require custom prompts and scanners.

## Model Access and Refusal Handling

deepsec ships with a classifier that checks whether a task was refused after each research step. According to the blog post, for the prompts deepsec uses, refusals are a non-issue for both Opus 4.7 and GPT 5.5 with off-the-shelf models. Both Anthropic and OpenAI offer "cyber" fine-tuned model variants for security tasks, and deepsec is compatible with those as well. For real scans, Vercel AI Gateway is recommended — one key covers both Claude and Codex, and the gateway's default quotas are sized for highly concurrent research.

## Update: Open-Sourced May 2026

deepsec was open-sourced on May 4, 2026, with the repository hosted at `vercel-labs/deepsec` on GitHub. The project is actively maintained, with the last push recorded in late June 2026. Getting started requires running `npx deepsec init` at the root of the target repository, which creates a `.deepsec/` directory for configuration and investigation catalogs.

## Features
- Agent-powered vulnerability scanning
- Regex-based candidate identification (scan stage)
- AI investigation with Claude Opus 4.7 and GPT 5.5
- Revalidation step to reduce false positives
- Git metadata enrichment for contributor identification
- Export findings as markdown or JSON
- PR/diff mode for CI gating
- Distributed execution via Vercel Sandbox microVMs
- Plugin system for custom regex matchers
- Resume interrupted scans automatically
- Works with existing Claude/Codex subscriptions
- Vercel AI Gateway integration
- Custom scanner plugins for auth models and data layers

## Integrations
Claude (Anthropic), Codex (OpenAI), Vercel AI Gateway, Vercel Sandbox, GitHub (git metadata), npm

## Platforms
WINDOWS, WEB, API, CLI

## Pricing
Open Source

## Links
- Website: https://github.com/vercel-labs/deepsec
- Documentation: https://github.com/vercel-labs/deepsec#docs
- Repository: https://github.com/vercel-labs/deepsec
- EveryDev.ai: https://www.everydev.ai/tools/deepsec