The Aikido Machine: on-prem AI pentesting that never leaves your network

Aikido Security’s new offering replaces cloud-hosted AI pentesting with a self-contained appliance that performs inference on local GPUs. This shift addresses regulatory mandates that prohibit moving proprietary code or repository data outside protected environments. By keeping the entire AI stack inside the customer’s network, the product sidesteps the compliance hurdles that previously blocked continuous AI testing for highly regulated entities.

The appliance is delivered as a pre-configured GPU server that Aikido engineers install and connect to power and network on the customer site. Once hooked up, the system can launch AI-driven scans continuously, eliminating the need to schedule discrete engagements or manage token-based usage limits. The on-prem design also means the organization retains full visibility into the models, prompts, and results, which is impossible with a black-box cloud service.

From an operational perspective, the machine provides built-in authentication, session handling, proxying, and scope enforcement, allowing it to exercise real login flows and stay within defined target boundaries. Each vulnerability discovered is accompanied by an executable exploit and a ready-to-merge code change, giving development teams concrete remediation steps without additional manual analysis. Early adopters, such as a major European bank, have already integrated the appliance into their security workflow.

Adopting the solution entails purchasing or leasing a GPU-equipped server and entering a service agreement for installation, updates, and support. Organizations must allocate rack space, power, and network connectivity, and they may need to plan for periodic hardware refreshes to keep pace with model performance requirements. The cost model is therefore centered on capital expenditure for the appliance and ongoing maintenance rather than per-test cloud fees.

The appliance’s effectiveness is limited to assets that reside within the protected network; code stored in external cloud repositories or services that require internet access cannot be scanned directly. Additionally, environments lacking sufficient GPU capacity or those that cannot accommodate the physical server will be unable to run the AI models at full speed. Finally, while the system can be updated manually, any need for rapid model upgrades may be constrained by the air-gapped design.