Security testing of LLM applications, RAG pipelines, AI agents, MCP servers and the AI model supply chain, against the OWASP GenAI Top 10 lists and MITRE ATLAS.
A full test of an application built on a large language model against the OWASP LLM Top 10 2026: prompt injection, sensitive information disclosure, output handling, excessive agency and abuse of cost, plus the usual web and API layers around it.
Testing retrieval-augmented generation: document ingestion and indirect prompt injection, vector store access control and tenant isolation, embedding weaknesses and leakage of retrieved content to users who should not see it.
Testing agents that call tools and act on users' behalf, against the OWASP Top 10 for Agentic Applications 2026: goal hijack, tool misuse, identity and privilege abuse, memory poisoning and unsafe code execution, with tests run in a controlled environment.
Testing Model Context Protocol servers and clients against the MCP security best practices: authorization and token audience, token passthrough, confused deputy, SSRF during OAuth discovery, session handling, scope design and local server sandboxing.
Review of where models, adapters, datasets and AI libraries come from and how they are verified, stored and updated, covering LLM04 Supply Chain and LLM05 Data and Model Poisoning; a review engagement rather than a live attack test.
A focused engagement on direct and indirect prompt injection, the top risk in the OWASP LLM Top 10 2026: how testers build scenarios from documents, web content and tool outputs, and how results show which guardrails held.
Objective-based adversarial testing of an AI system using MITRE ATLAS techniques and agreed harm scenarios, as opposed to a checklist test; covers what the EU AI Act's adversarial testing duties mean for buyers and how findings are reported.
Tell us what you're working on, whether it's a deployment, an audit, a security test or a cyber range. You'll speak with an engineer who can help you scope it.