IMPLICATOR
One repo this week instead of five. Vercel's deepsec turns coding agents loose on a codebase and asks them to find vulnerabilities the way a security engineer would. It cleared 7,754 stars. The more revealing document is the benchmark Vercel published alongside it, which shows the best model finding under a third of known bugs.
deepsec
An agent-powered vulnerability scanner built to review all the code in a large existing repo, not just the diff in a pull request. A regex pass narrows the field, then coding agents investigate each candidate file, trace data flow, check for mitigations already in place, and write up findings with severity ratings. Runs on your own infrastructure.
How it works
deepsec runs in five stages. A regex pass flags security-sensitive files without calling a model, which is why a 2,000-file project clears in about 15 seconds. Agents then investigate each flagged file, tracing data flow and checking whether a mitigation already exists. A second agent pass revalidates findings and adjusts severity. A fourth stage reads git metadata to name whoever last touched the code, and the fifth exports findings as tickets.
Setup runs through one command, and the spending limits are the part to read first:
# guided setup: pick a model, set a budget, start scanning
npx deepsec init --max-cost-usd 100 --max-duration 2h
pnpm deepsec scan # regex pass, no model calls, no cost
pnpm deepsec process # the agent investigation, where the money goes
pnpm deepsec revalidate # second pass to cut false positives
pnpm deepsec export --format md-dir --out ./findings
Interrupt a run and the next one resumes where it stopped, skipping files already scanned. Large repositories distribute across Vercel Sandbox microVMs, with API keys held host-side and injected only at egress, so a compromised agent never reads them. It works with a Claude or Codex subscription you already pay for.
The AI workflows worth keeping
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 4:45am PST. FREE!
No spam. Unsubscribe anytime.
The number Vercel published on itself
DeepsecBench arrived on July 27 and is the more useful artifact. It holds an open-source codebase at a commit from before its major security fixes landed, then scores models against 231 human-verified findings across 50 entry-point files, weighting recall twice as heavily as precision.
The best run found 30.7 percent of them. Twenty of the 25 runs came in under 20 percent. OpenAI's top reasoning tier led the table with a composite score of 35.58, at $55.98 and three hours 39 minutes per run. Claude scored 28.36 in 47 minutes for $31.96, a better trade for most teams on both clock and cost.
One absence carries its own information. Anthropic's most capable model does not appear in the table at all. Vercel's note says it declines security work, including defensive tasks.
Vercel puts the false-positive rate at 10 to 20 percent, which is what the revalidate stage exists to pull down. Scans of large codebases can cost thousands or tens of thousands of dollars.
| Assessment | What it means in practice |
|---|---|
| Advantages | Reviews an entire existing codebase rather than the diff, so legacy code written before current models gets a first look. Runs on your own infrastructure with no source shipped to a vendor. Resumable after interruption. Apache-2.0, with pluggable agents and a --max-cost-usd ceiling. Scales to 1,000+ concurrent sandboxes. |
| Disadvantages | Best-in-class recall is 30.7 percent on Vercel's own benchmark, with 20 of 25 runs under 20 percent. A 10 to 20 percent false-positive rate survives revalidation. Large scans run into thousands to tens of thousands of dollars.Not tuned for libraries or frameworks without customization. |
| Best use cases | An established application or service monorepo that has never had a line-by-line review. A pre-audit sweep to hand a security contractor a prioritized file list. Triage after inheriting a codebase. A one-time sweep of an acquisition target.Poor fit as a per-PR gate, or as your only control. |
Defense and offense read the same benchmark
The README's framing is the honest one: treat deepsec like a coding agent with full shell access, and stay alert to prompt injection arriving through external dependencies.
Twelve days ago an offensive-security skill pack topped GitHub Trending with more than 20,000 stars. deepsec is broadly the same capability aimed the other way, shipped by a company whose CTO wrote 89 of its roughly 94 commits himself. The benchmark is where the asymmetry shows. A defender running deepsec at current model quality catches at most a third of what is in the code. An attacker running the equivalent needs one finding to matter.
Two customers vouched for it in Vercel's launch post, which is where testimonials tend to live. "deepsec's scans have been the most thorough, with most findings, and good true-positive rate," said James Perkins, CEO of Unkey. Steven Tey, who founded dub.co, said it was "the first tool that's surfaced the kind of issues we'd actually want a security engineer to flag." Both ran it on open-source repos Vercel selected.
deepsec
Most security tooling built on models this year has been sold on the finding it produces. deepsec ships with a benchmark that says how much it misses, and the honesty is the reason to look at it. A 30.7 percent ceiling reframes the purchase: this is not a replacement for a security engineer, it is a way to point one at the right 200 files in a repo nobody has read since 2023.
Test it on a disposable clone of one service, not the monorepo. Set --max-cost-usd low enough that a runaway scan is an annoyance rather than an incident, run scan and process, then read the revalidate output rather than the raw findings. Success is not a clean report. It is two or three issues your existing pipeline never flagged, traced to files a human can now go read.
Frequently Asked Questions
Why only one repo this week?
Repo Radar normally runs five cards. This issue trades breadth for depth because deepsec shipped with a benchmark measuring its own limits, and that number deserved the room.
Are stars enough?
No. Stars measure attention. deepsec cleared 7,754 of them, and the figure that decides whether it is useful is the 30.7 percent recall in Vercel's own benchmark.
What does the difficulty score mean?
It estimates how hard the project is to test or adapt, not how impressive the underlying engineering is. deepsec rates 3 of 5 because setup is straightforward and the real barrier is spend.
What does a scan actually cost?
Vercel says scans of large codebases can run into thousands or tens of thousands of dollars. The init command accepts a --max-cost-usd ceiling, and the regex scan stage makes no model calls at all.
What should teams check before production use?
License, data retention, credential access, update speed, maintainer responsiveness, and whether the repo has a realistic rollback path.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Related stories
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
