Vercel deepsec is an open-source AI security harness. It sends coding agents through your codebase to trace data flows and rank vulnerability findings. It is a strong agentic code reviewer. It is still a static analyzer, so it can't test deployed authentication, business logic or real exploitability. Pair it with runtime penetration testing.
Vercel open-sourced deepsec on May 4, 2026. It is a security harness that runs coding agents over your codebase to find vulnerabilities. It launched on Claude Opus 4.7 and GPT 5.5. Our short answer: it handles half the problem well, and you should pair it with runtime testing for the other half.
This post covers what Vercel deepsec is and how its harness works. It then looks at its benchmark scores, where it falls short and why a runtime platform like MindFort belongs next to it.
What is Vercel deepsec?
Vercel deepsec is an open-source AI security harness that uses coding agents to find vulnerabilities in your codebase. It runs on your laptop or in CI. You pick the model, and findings come out ranked by severity. Vercel built it to scan its own monorepos and released it under the Apache 2.0 license .
Think of it as agentic SAST. Rule-based tools like CodeQL and Semgrep match code against a fixed library of patterns. With deepsec, LLM agents read security-sensitive files and trace data flows across the repo instead.
How does the deepsec harness work?
A run starts with npx deepsec init, according to Vercel's architecture docs . A setup agent studies the repo and writes an INFO.md file describing it. Next, a free regex scan flags security-sensitive files in about 15 seconds per 2,000 files. Coding agents then investigate each flagged file and return findings with severity ratings.
A second agent pass called revalidate re-checks those findings to cut false positives. The enrich step adds git blame so each issue has an owner. Export turns the findings into tickets for humans or coding agents. Agents run on Codex, Claude Code or the open-source Pi coding agent, and large repos can fan out across Vercel Sandbox microVMs.
How does deepsec perform on benchmarks?
Vercel publishes DeepsecBench , a leaderboard of models running inside the deepsec harness. It scores how well each model finds vulnerabilities in application code by combining recall and precision. As of October 30, 2026, the top setup is GPT-6 Sol at xhigh reasoning with a score of 40.75. It found 35.6 percent of the benchmark's vulnerabilities at 96 percent precision for $12.76.
No setup on the board finds more than 36 percent of the vulnerabilities. The top six keep precision above 90 percent, so their findings are usually real. Vercel runs the benchmark itself. Read it as a comparison of models inside deepsec.
Is Vercel deepsec good for cybersecurity?
For code review, yes. Vercel reports a 10 to 20 percent false positive rate on its own monorepos in the launch post . The deepsec FAQ now puts it at roughly 10 to 29 percent for high-severity findings after revalidation. Dub founder Steven Tey called it the first automated tool to surface the issues a security engineer would flag.
A few things make deepsec useful:
- It runs on your infrastructure. No third-party SaaS sees your source. You can run it with your existing Claude or Codex subscription.
- It scales out. Scans on Vercel's own codebases routinely reach more than 1,000 concurrent sandboxes.
- Its findings go deeper than legacy SAST. Rule libraries miss novel issues. In the IRIS study , LLM-augmented static analysis caught 69 of 120 real Java CVEs. The best traditional SAST tool caught 27.
What are the limitations of Vercel deepsec?
Under the hood, deepsec is still an LLM-powered static analyzer. Static analysis has a ceiling that no model upgrade will fix. There are three reasons.
1. Not all vulnerabilities live in code. Auth flows, rate limits, IAM policies and network segmentation only show up in a running system. So does business logic that spans several services. Cybersecurity Dive makes the point plainly. AI can say a code pattern might allow SQL injection, but only the running app shows whether it does.
2. The model that wrote the bug may be the one grading it. When one model family writes and audits the code, its blind spots repeat. The model is likely to overlook the mistakes it just made. There is no oracle. You only get another opinion.
3. Static findings stay theoretical until someone exploits them. A 10 to 20 percent false positive rate beats legacy SAST, where NIST measured rates up to 78 percent on Java . You still don't know which findings are exploitable in your deployment. As StackHawk notes , only runtime testing can confirm that.
Runtime testing closes this gap. It sends real requests, watches the responses and proves exploitation against the live system.
Here is how false positive rates compare across approaches:
| Tool / approach | Reported false positive rate | Source |
|---|---|---|
| Legacy SAST (Java) | Up to 78% | NIST, via Contrast |
| Legacy DAST (OWASP Benchmark) | 82% | Contrast |
| Vercel deepsec (agentic SAST) | 10 to 20% | Vercel |
| Modern signature-based DAST | 5 to 8% | Snyk |
| MindFort (validation-first runtime) | Below the 5 to 8% DAST floor | MindFort's own assessments |
Should you pair deepsec with a runtime testing tool?
Yes. Use deepsec for breadth across large codebases. Then run an autonomous runtime platform against the deployed app to confirm what is exploitable. It also catches IDORs, business logic flaws and chained exploits that never appear in source.
Comparisons of SAST and DAST find the two complementary. Traditional DAST has its own problem, which is noise. The OWASP Benchmark put legacy DAST false positives at 82 percent , and modern vendors sit around 5 to 8 percent . Agent-based runtime testing with codebase context closes that last gap.
What is the best alternative to deepsec for runtime security?
MindFort replaces traditional DAST. Our autonomous agents run continuous pentesting and exploit validation against your live apps, APIs and infrastructure. They run on MF-1, a purpose-built offensive security model. Every finding is validated through real exploitation, so false positives stay below the 5 to 8 percent DAST floor in MindFort's own assessments.
Codebase context is rolling out soon. MindFort will then run whitebox against your source while it attacks the running system. That covers the static side too. Book a demo .
FAQ
What is Vercel deepsec?
Vercel deepsec is an open-source AI security harness. It uses coding agents to investigate codebases and trace security-sensitive data flows. It then produces ranked vulnerability findings.
How does the deepsec harness work?
A free regex scan flags security-sensitive files. Coding agents investigate each one and report findings with severity ratings. A revalidate pass re-checks those findings to cut false positives before export.
Is deepsec good for cybersecurity?
Yes, for code review. It surfaces prioritized findings across large codebases. It does not replace runtime security testing or penetration testing.
What are deepsec's limitations?
It is a static analyzer. It cannot fully test deployed authentication flows, business logic or infrastructure configuration. It also can't prove that a finding is exploitable in the running app.
Should teams pair deepsec with runtime testing?
Yes. Pair static tools like deepsec with runtime testing. Runtime testing sends real requests, observes real responses and confirms what is exploitable.
About the author

Akul Gupta
Co-Founder & CTO · MindFort
AI researcher focusing on LLMs in cybersecurity. Red-teamed models for OpenAI and Anthropic as part of their safety programs. Published multiple conference papers. M.S. Computer Science, UIUC.