Skip to content
ForgeRheono · planned · model hardening

Planned · early access open

Train the model your team runs against a real attacker.

Your company runs its own AI model, and that model writes code. Forge trains it against Rheono Siege, our attacker, so the code it writes holds up when someone tries to break in. The hardened model stays on your own servers.

Request early access

Four questions. A person reads every request; a small design-partner group comes first.

Trains
The model you run
Graded by
A live attacker
Runs on
Your servers
Status
Planned
The forgeillustration · scroll to run it

One practice task, from the first line of code to your servers.

01 / 06 · Writeillustration
  1. 01

    Write

    Your model writes an app for a practice task, drawn from the kinds of code your team ships. Each tower is one endpoint.

  2. 02

    Test

    The tests run. Code that doesn’t do what the task asked stops here and earns nothing.

  3. 03

    Attack

    Siege attacks every endpoint that works. A break-in counts only when three checks agree: one AI finds it, a second fails to prove it wrong, and a plain script repeats it.

  4. 04

    Reward

    Only code that works and holds earns a reward. Standard training would pay for the red towers too. Forge pays nothing for them.

  5. 05

    Harden

    The reward moves the model’s weights. The next app it writes faces the same attack, and the aim is that fewer break-ins land.

  6. 06

    Deploy

    The hardened model runs on your own servers, the same way you run it today.

Illustration of one training step, not a measurement. The model writes an app of seven endpoints; one fails its tests; Siege breaks into three of the six that work; only the three that held earn a reward; the weights step; the next app takes the same attack; the hardened model goes to your servers.

Why AI code breaksSusVibes · per 100 tasks

Most AI-written code passes its tests and can still be broken into.

A coding model learns by practice: it writes code, the tests run, and it earns a reward when they pass. Nothing in that practice tries to break in, so the model learns to write code that works and can still be broken into.

An AI coder on real tasksone cell = one task
Of 100 real coding tasks, an AI coding agent got 12 working and secure, 45 working but breakable, and 43 not working.
Worked

57%

of tasks passed their tests.

Worked and secure

11.8%

also held when someone tried to break in.

Source: SusVibes, arXiv 2512.03262
One task in detailsample run

The model earns its reward only if the code works and Siege can’t get in.

Training runsample run
one practice task per row · newest last
  • 0417sign-up APItests passSQL injectionreward 0
  • 0418file uploadtests passpath traversalreward 0
  • 0419password resettests failnot attackedreward 0
  • 0420invoice APItests passno way inreward 1
  • 0421admin exporttests passmissing access checkreward 0
  • 0422login formtests passno way inreward 1

Reward 1 only for code that passes its tests and that Siege can’t get into.

Reward ruleper practice task
The reward for one practice task. Tests pass and Siege cannot get in: reward 1. Tests pass and Siege gets in: reward 0, where standard training pays 1. Tests fail: reward 0.
Siege can’t get inSiege gets in
Tests pass

1Reward

Works and holds. The only code that pays.

0No reward

Standard training pays 1 here.

Tests fail

0No reward

Doesn’t work.

0No reward

Doesn’t work.

Three checks keep the reward hard to fool. A model in training learns to fool whatever grades it, so the grade has to be a real break-in.

How we measure itresearch in progress

One number decides whether Forge works.

The share of answers that pass their tests and that no known exploit gets into, on benchmark tasks the model never trained on. We will publish the result whichever way it comes out.

  1. 01Held-out tasksScored on a public benchmark of web-app tasks, never the tasks the model practised on.
  2. 02A real controlThe same model trained on tests alone. Forge has to beat that, not just the untrained model.
  3. 03Fixed in advanceThe success bar, the seeds and the statistics are written down before training starts.
Who it is for

Teams that run their own model.

  • Banks and fintechs

    Strict data rules, and some already run their own models.

  • Software teams

    Moving developers to open coding models to cut token bills.

  • Regulated firms

    AI and data kept in their own region, on their own servers.

  • Siege customers

    The holes Siege finds in your apps become training tasks for your model.

Early accessdesign partners first

Request early access.

Tell us which open model you run and where. We pick a small first group of design partners whose setups we can train and measure well.

Design partners get
  • A hardened copy of the open model you already run, on your servers.
  • Before and after security scores on the kinds of code your team writes.
  • A say in what Forge measures and how it is priced.

Forge will come in two tiers, base model and premium model. Prices are set with the first design partners.

Requestearly access

A person reads every request. We are choosing a small design-partner group first.

Questions
Can I buy Forge today?
Not yet. Forge is planned, and early access is open. Siege, the attacker Forge trains against, is on sale today.
Which models does it work with?
Open-weight models your team already runs. Name yours in the request and we will say whether it fits the first group.
Where does the hardened model run?
On your own servers, the same way you run your model today. How training handles your code is agreed with each design partner before anything starts.
Does it work yet?
It is research in progress. We will publish the measured result, good or bad, before Forge goes on sale.
How is this different from a pentest?
A pentest finds holes in an app after it is written. Forge changes the model that writes the code, so fewer holes get written. Siege does the first; Forge uses Siege to do the second.
Rheono Forge

Code that survivesan attacker.

Tell us which open model you run and what it writes for your team. That is enough to start.

PostHog analytics (EU), only if you agree. Privacy