Adversarial Evaluation
Targeted adversarial scenarios for RAG systems and AI agents — prompt injection, manipulated context, unsafe tool requests and sensitive-information exposure.
The Challenge
Standard functional tests confirm that a system handles expected inputs. Adversarial evaluation tests what happens when inputs are deliberately crafted to cause unsafe, unintended or harmful behaviour.
We add scoped adversarial scenarios to an evaluation engagement to test how the system responds to manipulated instructions, context, documents, tool requests and attempts to expose sensitive information.
What Can Go Wrong
Adversarial categories we test
Direct Prompt Injection
Crafted input that overrides system instructions or bypasses guardrails.
Indirect Prompt Injection
Malicious instructions embedded in retrieved documents or context.
Manipulated Context
Poisoned or altered context that steers generation toward incorrect or harmful outputs.
System-Prompt Extraction
Attempts to reveal system prompts, configuration or internal instructions.
Sensitive-Information Exposure
Triggering the system to reveal personal, confidential or restricted data.
Unsafe Tool Requests
Crafted prompts that attempt to trigger unauthorised or dangerous tool calls.
Excessive Agency
The system takes actions beyond its intended scope or authority.
Confirmation Bypass
Attempts to skip required confirmation steps or claim false success.
Our Testing Methodology
How we approach adversarial evaluation
Scope Attack Surface
Identify which adversarial categories are relevant to the system and engagement.
Generate Scenarios
Automated attack generation provides coverage across identified categories.
Validate Results
Manual validation determines whether a reported result is a reproducible system failure.
Reusable Tests
Confirmed findings are converted into reusable tests that verify the problem does not return.
What You Get
Adversarial Report
Confirmed adversarial findings with evidence, severity and reproduction steps.
Attack Scenario Set
Categorised adversarial scenarios used during the evaluation.
Reusable Quality Test Suite
Reusable tests derived from confirmed findings, rerun after each change.
Add Adversarial Scenarios to Your Evaluation
Adversarial evaluation is not a penetration test, source-code security audit, compliance audit or certification.
Discuss Adversarial Scenarios