Our research
We publish what we learn about this attack surface.
Our work on AI tooling goes into public checklists and vulnerability write-ups, and straight back into how we test and build.
Industries · 06
LLM features, MCP servers, agents and developer platforms. We test what happens when untrusted text can reach tools, data and infrastructure, and help you build it safely.
Overview
AI features introduce a new class of vulnerability: prompt injection, data leaking through context, agents acting on behalf of the wrong user, and tool servers that trust whatever the model sends. They sit on top of the usual risks in APIs, authentication and cloud.
We research this area and publish what we find, including our MCP server security checklist and our write-up of the Ollama vulnerability known as Bleeding Llama. That research goes straight into how we test and build.
Frameworks we work to
Risks
What our testers look for first in ai & developer tools, from experience rather than a generic checklist.
Instructions hidden in documents, emails or web pages that the model follows.
Tool servers with broad permissions, no authentication or no checks on what the model asks for.
One user's data, system prompts or secrets surfacing in another user's answer.
Agents acting with their own permissions instead of the user's.
Inference servers and model APIs exposed to the internet without authentication.
Models, packages and plugins pulled in without review.
How we help
The services ai & developer tools clients use most. Each one has its own page with scope, process and FAQs.
LLM features tested for prompt injection, data leakage and authorisation alongside the usual web risks.
TestMCP servers, tool endpoints and model APIs tested like the attack surface they are.
TestGPU hosts, inference servers and the cloud accounts behind them.
BuildAI features and MCP servers built with least privilege and human approval where it matters.
Evidence
The requests that usually trigger a test in this sector, and the evidence that answers them.
NoteThis field moves quickly. We test against current public guidance and our own research, and say so when something is unproven.
Our research
Our work on AI tooling goes into public checklists and vulnerability write-ups, and straight back into how we test and build.
An attack where text the model reads, such as a document, email or web page, contains instructions that override what you intended. It matters most when the model can call tools or see other users' data.
Yes. We test authentication, authorisation, input handling and what each tool can actually do, and we have published an MCP server security checklist.
Yes. We test your application, prompts, tools and data flows. Vulnerabilities in the model provider's own service are their responsibility, but we can coordinate disclosure if we find one.
Yes. Our developers build LLM features and MCP servers with least-privilege tools, user-scoped access and human approval for risky actions.
A security engineer or developer reads every enquiry and replies, not a salesperson.