AI security
Secure the AI you build, use and face
We test AI systems the way attackers do, prepare your people for AI driven fraud, and put governance in place you can prove to a board or regulator.
AI assessment scope
Available now- Chat assistants and copilotsDirect and indirect prompt injection
- RAG pipelinesPoisoned documents and data leakage
- Agents and MCP toolsTool hijacking and excessive permissions
- Models and datasetsSupply chain and unsafe model files
Findings mapped to OWASP Top 10 for LLM Applications and MITRE ATLAS
Risks we test for
These are the risks we test for and prepare teams against. Each one maps to recognised frameworks.
- 01
Prompt injection
Instructions hidden in user input, web pages, emails or files that make a model ignore its rules. Indirect injection through content an agent reads is the harder case.
- 02
Agent and tool hijacking
An agent with access to email, code, payments or MCP tools can be steered into real actions for an attacker. Too many permissions turn a text bug into a breach.
- 03
Sensitive data leakage
Secrets, customer data or internal documents exposed through prompts, retrieval results, logs or model outputs.
- 04
Model and data supply chain
Poisoned training data, tampered models, and model files that run code when loaded. Third party models need the same checks as any other dependency.
- 05
Deepfakes and voice clones
Synthetic audio and video used to impersonate executives, family members or officials to push urgent payments or data requests.
- 06
AI scaled phishing and fraud
Fluent, personalised scam messages produced at volume in any language, without the spelling mistakes people were taught to spot.
- 07
Shadow AI
Staff pasting confidential data into unapproved AI tools and browser extensions, outside any contract, retention rule or audit trail.
- 08
Insecure output handling
Applications that trust model output and pass it straight into code, queries, HTML or system commands.
How we approach AI security
Area 1
Secure the AI you build
LLM apps and AI agents open new attack paths: prompt injection, tool hijacking and data leaks. We test them like an attacker, show the evidence, and help your team fix what we find.
- AI red teaming and LLM penetration testing
- AI agent and MCP tool security review
- Model and dataset supply chain review
Area 2
Defend against AI driven attacks
Attackers now use AI to write convincing phishing, clone voices and faces, and find weaknesses faster. We prepare your people and processes to recognise these attacks and stop them.
- Deepfake voice and video fraud readiness
- AI written phishing and email compromise
- Free scam checks for the public with Guardian
Area 3
Govern the AI you use
Most organisations already use AI tools nobody approved. We map where AI is used and what data it touches, then put clear rules and controls in place that people follow.
- Shadow AI discovery and data leak risk
- AI policy and acceptable use rules
- ISO/IEC 42001 and NIST AI RMF readiness
Findings your auditors already understand
We map every finding and control to public frameworks, so reports slot straight into your risk register. We reference these frameworks; we are not certified by them.
| Framework | How we use it |
|---|---|
| OWASP Top 10 for LLM Applications | Test coverage for LLM and agent applications |
| MITRE ATLAS | Adversary tactics and techniques against AI systems |
| NIST AI Risk Management Framework | AI risk governance and controls |
| ISO/IEC 42001 | AI management system requirements |
Launching an AI feature or agent?
Have it tested before it reaches customers or sensitive data. We scope an assessment in one call.
Under attack right now?
Ransomware, account takeover or fraud in progress. Email us with URGENT in the subject and it is read before anything else.
Get incident response