exec() in /convert
Reached from a route that takes a filename straight off the query string. The fix is written and waiting on your branch.
They exploit it
They fix it
Arvion is your autonomous product security team. They prove a vulnerability by walking the attack end to end, then fix it where it lives. They patch the code, fix your cloud infrastructure, or add guardrails on an over-permissioned agent.
Who you employ
Five security engineers, at your service
Each engineer is an autonomous agent with a specialized role. The team works together to safeguard your code.
Calvin
Calvin is an expert on finding anything related to AI features in your code. Calvin will search through your source code and identify agentic components, models, MCPs prompts, or chained tools, and then analyze them for potential security issues.
“You have more agents running than your list shows. I find them and flag the ones with no authentication.”
Live demo
See it run on a stack like yours
Pick a stack and press Run. Ten seconds, nothing to install — watch what gets ruled out, and what gets fixed.
Stack
acme/checkout-api · src/routes/refunds.jsSample
Wright · fix opened — PR #4127 closes the path. The merge is yours.
The queue
9.8lodash 4.17.20 · prototype pollutionFILED · the vulnerable _.merge path is never called
ALERTexec() in /healthz version handlerFILED · constant argv — no user input reaches it
ALERTPOST /refunds/:id/receipt● REACHABLE · orderRef flows into a shell string
The chain — severed
- POST /refunds/:id/receipt● proven
orderRef in the JSON body
- orderRef → shell string● proven
zero sanitizers on the path
- ✓exec() runs it via /bin/sh -cclosed
neutralised — src/routes/refunds.js
- RCE as the API service accountunreachable
the fix closed the path before it got here
The fix
const { orderRef } = req.body;
+ if (!/^[A-Za-z0-9_-]{1,32}$/.test(orderRef)) {
+ return res.status(400).json({ error: "bad ref" });
+ }
- const cmd =
- `receipt-gen --order ${orderRef} …`;
- exec(cmd, (err) => {
+ execFile("receipt-gen",
+ ["--order", orderRef, "--out", out],
+ (err) => {
PR #4127 — refunds: execFile + allowlist for receipt genYours to merge ▸
Filed as noise: the lodash 9.8 (unreachable) and exec() in /healthz (constant argv).
A specimen. Your own run is composed from one source you connect — a repository today; your code, dependencies, IaC and agent configs come with it.
Three findings, Only one worth fixing
Security tools find lots of potential issues, but not all of them are worth your time. Arvion tells you which findings are worth your time and why.
lodash 4.17.20
Real advisory, but the vulnerable path is never called from your code. Filed with the reason, so nobody re-opens it next quarter.
credentials in env
We could not decide this one either way. It is marked assumed rather than quietly counted as clean.
Book 30 minutes
Bring one repo. You leave the call knowing which of your findings were worth the afternoon.
The limits you set
Three laws an autonomous engineer follows
Written for machines that act, guarding the one thing a security team can break — your trust
- First Law
An engineer may not claim more than it proved — nor, by staying silent, let you believe you are safe when you are not.
a blast radius it can't prove is published as "ASSUMED".
Binds Saga and Lyra hardest because they are the ones that report their findings.
- Second Law
An engineer works to the clearance you set and no further — except where obeying would break the First Law — and hands you every call that is yours.
the PR opens, but a human needs to review.
Binds Wright alone, because Wright is the only one making any changes.
- Third Law
An engineer guards the one thing that lets it exist — your trust — but never by hiding a failure, faking a preview, booking a win it did not earn, or reaching past the clearance you set.
a scope it couldn't reach is displayed as "NOT COVERED".
Binds the whole team. If you disable an engineer, their scope is marked "NOT COVERED" rather than counted as clean.
Stop operating security software. Put the team on retainer.
Pick the domain that's loudest — your AI agents, your code, or everything you pull in. One engineer works it end to end: what is worth fixing, the fix written, delivered where your team already works. The rest gets filed, with the reason printed.
Book 30 minutes →A work call: you connect one source, we run it live, and you leave with whatever it found and wrote.
- One repo to start with read access
- No agent runs without your clearance
- You decide what gets fixed
Prefer email to calls?
Leave one source and we will come back with what it found and the fix it wrote.