AI pentesting entirely under your control

Nothing leaves: not your code, not your findings, not even telemetry. Aikido Machine pulls updates from a separate machine and everything else runs locally. If you'd rather have updates arrive automatically, you can whitelist a single update domain instead. In that mode Aikido Machine reaches out only to check for and pull updates on demand. Your code, repositories, and findings stay inside your network either way.

Everything runs in isolated containers, and every request the agents make passes through a proxy that checks it against the scope you defined. Data at rest is encrypted, and Aikido Machine sits inside your network under your own access controls, so it's governed like any other asset in your data center.

Because the model is only part of the picture. A real attacker works from the outside, guessing at how your application is built. Aikido Machine works with your source code, dependencies, and runtime in front of it, so it reasons about the app the way your own engineers do. That context finds real, exploitable flaws that a few extra points of raw model capability won't. Combined with a harness built specifically to drive offensive work, open-weight models running locally consistently match or beat what a frontier model manages blind.

Not without your say-so. The models are fixed to the version you're running, so nothing shifts underneath you mid-engagement. Our engineers test new model releases as they come out, and as soon as one outperforms the model you're on, we propose the change to you. Updates, including model changes, arrive on your schedule and only when you choose to apply them, so your results stay consistent from one pentest to the next.

Today Aikido Machine runs AI pentesting and AI Code Analysis. Capabilities that already run on-prem like static analysis, dependency, secrets, container, and virtual machine checks are planned by redirecting those existing tools. Two things won't run on Aikido Machine by nature: attack surface management and dynamic application security testing (DAST), because both work from outside your network looking in. Runtime protection ships as a software development kit (SDK) that goes into your own backend.

Yes. Because Aikido Machine works from your actual source code, dependencies, and runtime rather than guessing from the outside. The language coverage is broad thanks to the use of models that were trained on a large variety of ecosystems. Therefore, older stacks, including COBOL, are in scope.

You define the scope, and the agents stay inside it. Every request they make passes through a proxy on Aikido Machine that checks it against your scope and blocks anything outside of it. Testing runs non-destructively by default, and anything more aggressive, like pushing an exploit further to confirm it, is something you turn on deliberately. These are the same controls the cloud pentesting agents run under, so Aikido Machine doesn't add risk.

Two configurations are available, mostly depending on the size of your largest target codebase and the total number of lines of code (LOCs) that ought to be pentested in a year.

Aikido handles the on-site install with a forward deployed engineer who sets it up in your data center and works through your network restrictions with your team. We plan it with you and send prerequisites ahead of time, things like network access rules, ports, power, and cooling.

You lease Aikido Machine, which means that one annual fee covers the hardware, the software, upgrades, support, and the service level agreement (SLA). If a GPU fails, we replace it. There are no per-pentest fees and no token costs, because inference runs locally.