
An applied AI lab for security
Five AI engineers trace what an attacker can actually reach in your code, dependencies and AI agents, write the fix and open the pull request. Your team reviews and merges.
Who it's for
A few security people, dozens or hundreds of engineers, and AI writing more of the code every week. The findings pile up faster than anyone reads them, and most can't be used by an attacker. Arvion adds a team that finds the ones that can, writes the fix and opens the pull request, so your people spend their time on the decisions.
Try it · sample app
This refund page has a flaw our team found in a sample repo. Get the server to run a command of your own. If you’d rather skip the puzzle, one tap shows you the trick.
214findings in this sample repo. One of them starts at this field.
Server log
POST /refunds/42/receipt {"orderRef": "A-1093; id #"}
$ receipt-gen --order A-1093; id # --out /tmp/rcpt-1730812.pdf
receipt-gen: --out is required
uid=1001(api) gid=1001(api) groups=1001(api)
↑ id ran on the server as the api user
Same trick. Nothing ran.
SSaga · application code
I found this route in the code before you did.
A-1093; id #WWright · remediation
The fix is ready: PR #4127. It checks the order reference and runs receipt-gen without a shell. Tests pass.
const { orderRef } = req.body;+ if (!/^[A-Z]-\d{1,8}$/.test(orderRef))+ return res.status(400).json({ error: "bad ref" });- exec(`receipt-gen --order ${orderRef} --out ${out}`, done);+ execFile("receipt-gen", ["--order", orderRef, "--out", out], done); + it("rejects shell syntax in orderRef", () =>+ post({ orderRef: "A-1093; id #" }).expect(400));
Deployed as v1.8.3. Now send the same trick.
POST /refunds/42/receipt {"orderRef": "A-1093; id #"}
400 bad ref
Nothing ran. The check stopped it before any command was built.
What the team did on this repo
Three of the 213 it filed, each with its reason
Bring one data source. You leave the call knowing which of your findings were worth the afternoon.
Who you employ
Each engineer is an autonomous agent with a specialized role. The team works together to safeguard your code.

Saga is a tracing expert. She traces untrusted input from the entry through your source code to the places it could do damage, and reads configuration and infrastructure definitions for hardcoded credentials, unsafe settings, or anything opened to the whole internet.
“I traced how a request to POST /refunds can reach a shell command. Here is the exact path it takes.”
The limits you set
The same three rules apply to every engineer on the team. The product enforces them word for word.
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.
In practiceEvery step of an attack path is marked traced or assumed. Code the team couldn't analyze is reported as not analyzed.
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.
In practiceYou set what each engineer may do. After ten fixes of one kind are merged in a row, the team asks to open that kind on its own. If too many get rejected, it steps back. Nothing merges without your team.
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.
In practiceA finding the team sets aside comes with the reason written down. A run that didn't finish says so.
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 15 min →A work call: you connect one source, we run it live, and you leave with whatever it found and wrote.
Leave one source and we will come back with what it found and the fix it wrote.