What we test
Three risks tested. One layer of evidence around them.
Security, reliability, and fairness and inclusion are the risks we test. Governance sits around them as the evidence that you have tested them and addressed what was found.
Select an area to see what we check.
Risk tested
AI Security
Can the AI be manipulated, or made to expose information or take actions it shouldn't?
We test:
- Prompt injection, both typed and hidden in documents, emails or web pages the AI reads
- Jailbreaks
- Leakage of personal or confidential data
- Agents with more access than the job needs
- Poisoned source content
Governance only holds up if the other three have actually been tested. Without testing, it's just paperwork.
How we rank risk
We rank each risk by what your AI can do and what it can reach, not just how serious the flaw sounds technically.
Standard vulnerability scoring wasn't designed for AI: a read-only chatbot and an agent with access to your systems can share the same flaw but carry very different risks.
What it means for you
One assessment, read four ways.
Every finding in business terms.
We translate every technical finding into business terms: what it could cost you in liability, reputation or regulatory action, and what to prioritise first.
Adversarial testing built for AI, not adapted from traditional pen testing.
Traditional security testing assumes fixed code. AI behaviour depends on data, changes with every model update, and can be manipulated through language alone. We test for that.
How we test
- AI red teaming. We attack your AI the way a real adversary would, including direct and indirect prompt injection, jailbreaks, sensitive data extraction and attempts to misuse agent permissions.
- Threat modelling. Before testing, we map your AI's components and data flows and identify likely attack paths, extending STRIDE with AI-specific methods.
What we test for
- Prompt injection, typed directly or hidden in documents, emails or web pages the AI reads
- Sensitive data and knowledge-base leakage
- Poisoning of data sources and vector stores
- Excessive agency: agents with more tools, permissions or autonomy than the task needs
- Tool and integration abuse
- Model and supply chain risk: third-party models, libraries and APIs
- Runaway behaviour, such as uncontrolled loops and resource consumption
Frameworks we align to
- OWASP Top 10 for LLM Applications
- OWASP Top 10 for Agentic Applications
- MITRE ATLAS
- NIST AI Risk Management Framework
Tooling
We combine established open-source AI testing tools with manual adversarial testing, because automated scans alone miss the creative attacks real adversaries use.
Who tests
Testing is led by senior AI specialists working with a CREST-accredited security partner. Meet the specialists
Test results turned into evidence.
We turn test results into evidence: documented risks, how they were ranked and what was found, aligned to the NIST AI Risk Management Framework, ISO/IEC 42001 and the EU AI Act, and to the Equality Act 2010 for fairness.
Testing matched to your architecture.
A chatbot using your documents is most at risk of leaking or being fed false information. An agent acting in your systems is most at risk of being hijacked or doing more than it should. You get precise technical findings your engineers can reproduce and fix.
Board and executives
Every finding in business terms.
We translate every technical finding into business terms: what it could cost you in liability, reputation or regulatory action, and what to prioritise first.
Security leaders
Adversarial testing built for AI, not adapted from traditional pen testing.
Traditional security testing assumes fixed code. AI behaviour depends on data, changes with every model update, and can be manipulated through language alone. We test for that.
How we test
- AI red teaming. We attack your AI the way a real adversary would, including direct and indirect prompt injection, jailbreaks, sensitive data extraction and attempts to misuse agent permissions.
- Threat modelling. Before testing, we map your AI's components and data flows and identify likely attack paths, extending STRIDE with AI-specific methods.
What we test for
- Prompt injection, typed directly or hidden in documents, emails or web pages the AI reads
- Sensitive data and knowledge-base leakage
- Poisoning of data sources and vector stores
- Excessive agency: agents with more tools, permissions or autonomy than the task needs
- Tool and integration abuse
- Model and supply chain risk: third-party models, libraries and APIs
- Runaway behaviour, such as uncontrolled loops and resource consumption
Frameworks we align to
- OWASP Top 10 for LLM Applications
- OWASP Top 10 for Agentic Applications
- MITRE ATLAS
- NIST AI Risk Management Framework
Tooling
We combine established open-source AI testing tools with manual adversarial testing, because automated scans alone miss the creative attacks real adversaries use.
Who tests
Testing is led by senior AI specialists working with a CREST-accredited security partner. Meet the specialists
Risk, legal and compliance
Test results turned into evidence.
We turn test results into evidence: documented risks, how they were ranked and what was found, aligned to the NIST AI Risk Management Framework, ISO/IEC 42001 and the EU AI Act, and to the Equality Act 2010 for fairness.
Technical teams
Testing matched to your architecture.
A chatbot using your documents is most at risk of leaking or being fed false information. An agent acting in your systems is most at risk of being hijacked or doing more than it should. You get precise technical findings your engineers can reproduce and fix.
How it works
Five steps. Results in two to four weeks.
A repeatable method that finds what your AI gets wrong, shows you why, and leaves you with the tests to keep checking it yourself.
Step 1: Agree outcomes
We agree what “good” looks like for your use of AI and your risk appetite.
We meet the people who own the AI to agree its purpose, who it serves, what it can access and how much risk you're willing to carry. This sets the standard every finding is judged against.
Step 2: Benchmark
We compare your AI against known failure patterns from live AI systems.
We compare how your AI is set up and used against failure patterns we've seen in live AI systems, so testing focuses on where real systems actually break.
Step 3: Test
We test across all four areas and rank every finding by severity.
We test security, reliability, and fairness and inclusion, and review governance. Every finding is logged, reproducible and ranked by severity.
Step 4: Recommend
Clear, prioritised guidance your team or chosen provider can act on.
We present the scorecard and findings, with practical guidance your own team or chosen provider can act on. We don't sell the fixes.
Step 5: Re-test and prove
Once changes are made, we re-test to confirm they work, and hand over the test harness so you can verify results yourself.
Once changes are made, we re-run the tests to confirm they worked. Then we hand over the test harness so you can re-run every test after any future change, without us.
What you receive
- A scorecard for each of the four areas
- Every finding ranked by severity, with a clear recommendation
- Every finding written twice: a plain-English summary for your board, and precise technical guidance for whoever fixes it
- An evidence pack for your board, regulator and clients
- A prioritised roadmap
- The test harness itself, so you can re-run every test after any change, without us
What we need from you
A few hours from the people who own the AI, mainly at the start and when we present findings, plus access to the AI in scope, agreed before we start and under your controls.
Independent by design
We never assess systems we built, and we never sell fixes for what we find.