01
Could outside content redirect your AI?
Assess how prompts, documents and other untrusted content could influence AI behaviour, including prompt injection and attempts to override intended instructions.
AI & LLM application testing
Add AI features.
Understand the risks.
An AI feature can introduce new ways to access information or trigger actions. We assess the permissions and connected workflows around your AI application, giving you practical findings for the version you plan to release.
Private, self-hosted AI assistance. Assessment material is not sent to third-party model providers for that assistance.
Questions your test can answer
Test the controls.
Understand the implications.
The OWASP Top 10 for LLM Applications forms our baseline, with additional OWASP guidance for agentic systems where relevant. We build on it with specialist testing tailored to your models, data flows, tools and workflows.
02
Could your AI reveal sensitive information?
Examine what information AI features can access or disclose through responses and connected workflows, including boundaries between users and organisations.
03
Does retrieval respect access and trust boundaries?
Review knowledge sources, retrieval permissions and the integrity of retrieved content, including whether information reaches only its intended users and workflows.
04
Is generated output handled safely?
Assess how model output is displayed, processed or passed to other systems, and whether application controls prevent it from causing unintended actions.
05
Can agents and tools act beyond their intended role?
Examine permissions, approval controls and limits on connected actions, so you can understand where AI could affect data or business processes without the intended oversight.
06
Are the systems around your AI protected?
Review connected services, dependencies, configuration and usage limits, alongside the web and API controls supporting your AI features within the agreed scope.
Your assessment output
Evidence for your team.
Clarity for your reviewer.
Your report includes an executive summary, a summary of findings and detailed findings with descriptions, reproduction steps, remediation steps and references. A readout helps your team understand the priorities and next steps.
Your assessment report
Clear findings.
A useful next step.
Executive summary
Key findings & implications
Summary of findings
An overview of the issues
Finding details
Understand, reproduce & remediate
Each detailed finding includes a description, reproduction steps, remediation steps and references.
Getting started
Bring the context.
We’ll shape the scope.
Bring a simple architecture diagram, the model and provider arrangement, representative workflows, permitted test data and the connected tools that may be assessed.
Two different questions
The AI in your product.
The AI we use to assist.
01
Your application
Your product may have its own models, providers and integrations. We assess these only within the authorised application scope.
02
Our delivery assistance
Private, self-hosted AI assistance. Assessment material is not sent to third-party model providers for that assistance.
03
The full arrangement
Self-hosting alone does not establish residency, training settings, logging or retention. These details need to be agreed for the engagement.
Let’s talk through
your penetration test.
Tell us about your product, your deadline and the evidence your reviewer needs. We’ll agree the scope and handling arrangements before work starts.