AI & LLM penetration testing

AI systems move fast. We test what they can do.

Independent testing of LLM applications, AI agents and connected tools — finding where untrusted input can cross into private data, privileged actions or unsafe output.

TEST EVERY TRUST BOUNDARY
INPUTUntrusted content
→
MODELRetrieved data
→
ACTIONTools & permissions
Business impactWhat the system can actually be made to do
Choose the right test

Four ways to test AI risk.

01

LLM application penetration test

Prompt injection, data leakage and unsafe output across the full application path.

02

Agent & MCP security review

Tool permissions, excessive agency and trust boundaries in autonomous systems.

03

AI red team

Objective-led adversarial testing of AI-enabled services.

04

AI governance assessment

Controls, oversight and policy for responsible AI adoption.

Show the work

A method you can inspect.

A risk-led process. Scenarios are adapted to your architecture and mapped to the OWASP LLM Top 10 (2026) and MITRE ATLAS. Every test is authorised before it runs.

Scope & threat model

Understand use cases, data and what matters.

Map trust boundaries

Where untrusted input meets data and actions.

Design scenarios

OWASP LLM, MITRE ATLAS and the OWASP Agentic Top 10.

Execute & verify

Approved test accounts and data only.

Report & retest

Evidence, fix priorities and revalidation.

RiskOWASP LLMMITRE ATLAS
Prompt injectionLLM01:2026AML.T0051
Sensitive information disclosureLLM02:2026AML.T0057
Excessive agency / tool misuseLLM03:2026AML.T0053
Unbounded consumptionLLM06:2026AML.T0034
Hidden context exposureLLM08:2026AML.T0056
RAG poisoningLLM09:2026AML.T0070
Fixed pricing
From £6,500 + VAT

Agent Security Test for an AI agent, copilot or LLM feature — typically two weeks from kick-off to report.

  • Agent Security Test£6,500
  • Multi-agent, RAG or MCP-connected systems£12,500
  • Quarterly retesting£1,800 a month

Prices exclude VAT.


  • One retest included in the price
  • An attestation letter you can share with customers and auditors
  • Tests the complete application path, not just the model
  • AI red teaming and governance assessments quoted to scope
FAQs

Answers before you scope.

How much does an AI security test cost?

An Agent Security Test is a fixed £6,500 + VAT. Multi-agent, RAG or MCP-connected systems are £12,500 + VAT. Quarterly retesting is available at £1,800 + VAT a month. One retest is included in the price.

How long does an AI test take?

Typically two weeks from kick-off to report, once access and test data are ready.

What do you need from us?

An architecture overview, use cases, test accounts, data sources, tools/MCP servers, permissions, environment details and an escalation contact.

Do you only test the model?

No — we test the complete application path: inputs, retrieval, tools, permissions and the actions the system can take.

Ready to test what your AI can do?

Share your architecture and we’ll propose a focused, fixed-price scope.