AI model pentesting

Understand how your model can fail.

Specialist delivery and availability are confirmed during scoping.

THE RIGHT STARTING POINT

For teams building, adapting, or operating machine-learning models. Feasibility depends on permissions, model type, representative data, and qualified specialist availability; an API-only test cannot establish training-data integrity.

A clear scope. A useful outcome.

Begin with the model’s task, deployment context, training process, and the information available to an assessor. Select meaningful adversarial tests rather than treating every model as a chatbot. Record datasets, model versions, query budgets, baselines, and limitations so your team can interpret the results. A hosted model endpoint permits different conclusions from access to weights, training data, or a controlled training pipeline.

The final scope, deliverables, and timing are agreed for your engagement.

What we cover

• Evasion and robustness scenarios relevant to the model’s task
• Privacy leakage, membership inference, and extraction risk where testable
• Poisoning and backdoor exposure in authorized data or training environments
• Model access controls, evaluation boundaries, and mitigation trade-offs

What you take away

• Threat model, access assumptions, and evaluation protocol
• Reproducible test results with baseline comparisons and limitations
• Prioritized mitigation options and residual-risk decisions
• Repeatable evaluation plan for future model or data changes

How to engage

Feasibility workshop, bounded model evaluation, and an agreed follow-up assessment.

What progress looks like

Coverage of agreed threat scenarios and measured change against the recorded evaluation baseline.

Before the engagement.

Agree the systems and business processes in scope, the authorisation required, the people involved and how the findings will be used. Timelines, deliverables, exclusions and any verification work should be confirmed directly.