Prompt injection: The defining vulnerability of LLM applications. We test both direct injection, where a user manipulates the model through crafted input, and indirect injection, where malicious instructions are hidden in content the model later processes, such as a document, web page, or email. This is where most serious AI compromises begin.

Jailbreaking and guardrail bypass: We probe the safety controls and system prompts meant to keep your model in line, testing how easily they can be overridden to produce prohibited output or unintended behaviour.

Sensitive data disclosure: We test whether your model can be led into revealing confidential information, including training data, system prompts, credentials, or the data of other users, a critical risk for any multi-tenant or data-connected application.

Insecure output handling: An LLM’s output is untrusted input to whatever consumes it. We test whether model responses can trigger downstream vulnerabilities such as cross-site scripting, SQL injection, or command execution in your wider application.

Excessive agency and agentic risks: As models gain the ability to call tools, access APIs, and take actions, the stakes rise sharply. We test AI agents for excessive permissions, unsafe tool use, and the risk of being manipulated into performing damaging actions on your behalf.

RAG and knowledge base security: For retrieval-augmented generation systems, we test the pipeline that feeds your model, assessing access controls, data poisoning risks, and whether users can reach information they shouldn’t through the model.

Model and API security: We test the infrastructure around your AI, including authentication, rate limiting, and the APIs that expose it, as well as risks such as model denial of service and model theft.

Supply chain and plugin security: We assess the third-party models, plugins, and components your system depends on, where a single insecure integration can undermine the whole application.

ISO 42001: independent testing to evidence the security controls within your AI management system, the first international standard dedicated to responsible AI.

OWASP Top 10 for LLM Applications: testing against the recognised benchmark for large language model security.

NIST AI Risk Management Framework: supporting a structured, evidence-based approach to AI risk.

EU AI Act: helping you demonstrate the risk management and robustness expectations placed on higher-risk AI systems.

ISO 27001, SOC 2 and GDPR: aligning your AI security with your wider information security and data protection obligations.

Executive summary: A plain-English overview of what we tested, what we found, and what it means for your business, written so a non-technical reader can understand your AI risk position in a couple of minutes.

Risk-prioritised findings: Every vulnerability we identify, rated by severity and real-world impact, so you know exactly what to fix first. No noise, no padding, just the issues that matter, in the order they matter.

Technical detail and proof of concept: For each finding, a clear explanation of the vulnerability, proof-of-concept evidence showing how it could be exploited, and everything your team needs to reproduce and verify the issue.

Practical remediation advice: Actionable, specific guidance on how to fix each finding, from strengthening system prompts and input handling to tightening the permissions of your AI agents. This is the part clients tell us they value most.

Compliance mapping: Where relevant, findings are mapped to the standard you’re testing against, including ISO 42001 and the OWASP Top 10 for LLM Applications, so the report slots straight into your governance or certification process.

Debrief and support: The report isn’t the end of the conversation. We walk your team through the findings in a debrief session, answer their questions, and stay on hand as you work through remediation.