Witcher
Services /

LLM Red Teaming

Challenge your LLM application with adversarial scenarios before those failures affect users or customers.

Discuss your testing needs

What the work covers.

For teams building or deploying LLM applications, copilots and RAG systems that need evidence of how their controls behave under attack.

LLM red teaming uses adversarial scenarios to investigate how an application can be pushed outside its intended behaviour. We test the configured system, including its instructions, retrieved content and integrations. The useful result is an account of the conditions that produced a failure, its consequence and the control that needs to change.

  • 01Direct prompt injection and jailbreaksAttempt to cross agreed instruction and behaviour boundaries through user inputs, including multi-step conversations.
  • 02Indirect prompt injectionPlace adversarial instructions in documents or other content the application reads, then examine how it handles the conflict.
  • 03Sensitive information disclosureProbe for unintended disclosure through responses and retrieval, using agreed test data and access boundaries.
  • 04Impact and repeatabilityRecord the application configuration, inputs and observed behaviour so teams can investigate a finding and assess its practical consequence.

Evidence your team
can act on.

An account of the adversarial scenarios tested, reproducible findings, observed impact and practical recommendations. We document the limits of the exercise and can retest agreed changes.

Scope, assessment criteria, test access and deliverables are agreed before work begins. See our assurance methodology.

Questions before
you start.

Does a successful exercise prove the model is safe?

No. Results apply to the application, configuration and scenarios tested. Model updates, new documents, different prompts or added tools can change its behaviour and may require further testing.

Can you test a RAG application?

Yes. Retrieval introduces its own questions about document access, hostile source content and disclosure. Those boundaries can be included in the agreed exercise.

Part of a wider
assurance engagement.

Explore how this work fits into AI Product Assurance, or compare all assurance services.

Testing for your
AI system.

Explore specific testing needs within an assurance engagement.

01

AI Security Testing

Understand how your AI application handles hostile inputs, sensitive data and the access it has been given.

02

LLM Red Teaming

Challenge your LLM application with adversarial scenarios before those failures affect users or customers.

03

AI Agent Security Testing

Test what your agent can read, change and trigger, and whether its permissions match the job it has been given.

Scope your AI testing.

Tell us about the system, the access it has and the decision the evidence needs to support.

Talk to Witcher