Skip to content

GenAI Application
Security Assessment.

Test what your AI can access, reveal and execute.
Manual adversarial testing across prompts, retrieval and connected tools.

Discuss your AI application
FOLLOW THE ATTACK. VERIFY THE IMPACT. Illustrative test path
The GenAI application attack surfaceUntrusted prompts and retrieved content reach orchestration, which interacts with the model. Tool calls must cross an authorization boundary before connected tools can reach business systems. Animated paths illustrate test inputs, not a measured security result. AI APPLICATION ADVERSARIAL INPUTS Prompt injection Retrieval abuse Tool misuse AUTHORIZATIONBOUNDARY RetrievalKnowledge & context OrchestrationInstructions & decisions ModelGeneration & reasoning Connected toolsActions & permissions Business systemsData · APIs · workflows TRACE THE INPUTINSPECT EXECUTIONVERIFY THE IMPACTThe GenAI application attack surfaceTest prompt injection, retrieval abuse and tool misuse. Trace inputs through retrieval, orchestration and the model. Inspect authorization before tool actions reach business systems. ADVERSARIAL INPUTSPrompt injection · Retrieval abuseTool misuse AI APPLICATION RetrievalKnowledge & context OrchestrationInstructions & decisions ModelGeneration & reasoning AUTHORIZATION Connected toolsActions & permissions Business systemsData · APIs · workflows
The model is one component.We test the controls connecting AI to your data and business operations.

WHERE SECURITY BOUNDARIES FAIL

A convincing answer.
An unauthorized outcome.

We test whether manipulated inputs become unauthorized disclosures, actions or downstream execution.

ATTACK SURFACEWHAT WE TESTOWASP LLM / 2025

Prompt injection

Direct prompts and instructions embedded in retrieved content that redirect application behavior.

LLM01Prompt Injection

Retrieval access

RAG retrieval that returns documents outside the requesting user's permitted corpus.

LLM08Vector and Embedding Weaknesses

Agent authority

Tool calls that exceed delegated permissions or bypass required approval for sensitive actions.

LLM06Excessive Agency

Output execution

Generated content passed into HTML, queries or commands without appropriate validation.

LLM05Improper Output Handling

Sensitive data

Secrets, personal data or restricted context disclosed through responses, memory or connected tools.

LLM02Sensitive Information Disclosure

Resource abuse

Missing token, request or agent-loop limits. Validation uses agreed budgets and stop conditions.

LLM10Unbounded Consumption

Selected test areas, not an exhaustive checklist. Supply-chain exposure, poisoning paths and deployment controls are included when relevant to scope.

ENGINEER-LED ADVERSARIAL ASSESSMENT

Follow the evidence.
Establish the impact.

A convincing model response is only the starting point. We inspect the application behavior behind it to determine what the attempted attack actually achieved.

THE EVIDENCE TRAIL

  1. 01

    Input & conversation state

    Record the test input, user role and relevant context.

  2. 02

    Retrieval & tool traces

    Inspect retrieved content, tool arguments and permission checks.

  3. 03

    Observed security impact

    Verify data exposure or unauthorized actions using agreed test data.

WHAT MAKES A FINDING

A reproducible path
to a verified outcome.

Findings connect the input to the affected control and observed impact. Each includes the conditions required, supporting evidence and practical remediation guidance.

Test conditionsExecution evidenceImpact & limitations
  1. 01 / DEFINE

    Agree the scenario

    Scope, identities, safeguards and stop conditions.

  2. 02 / TRACE

    Inspect execution

    Follow inputs through retrieval, model and tool calls.

  3. 03 / VALIDATE

    Confirm the impact

    Repeat relevant attempts and document observed outcomes.

  4. 04 / RETEST

    Verify the fix

    Repeat agreed cases and record the remediation status.

REFERENCED FRAMEWORKS

Standards with
a defined purpose.

We select references for the system being assessed and document the version, applicable controls and coverage gaps.

OWASP LLM Top 10 APPLICATION RISK TAXONOMY
Map relevant findings to the 2025 risk categories shown above.
MITRE ATLAS ADVERSARY BEHAVIOR
Inform threat hypotheses and relevant AI attack techniques.
NIST AI RMF RISK CONTEXT / AI 600-1
Use the Generative AI Profile to inform risk discussions, evaluation context and recommendations.
CIS Benchmarks SUPPORTING INFRASTRUCTURE
Reference applicable configuration guidance when cloud, container or host hardening is in scope.

Framework references guide the assessment; they do not constitute certification or a full AI governance audit.

START WITH YOUR ARCHITECTURE

Bring the system.
We'll define the test.

Scope a GenAI assessment

USEFUL FOR THE FIRST DISCUSSION

  • 01Application flow & model providers
  • 02Data sources & retrieval permissions
  • 03Agent tools & approval boundaries
  • 04Test environment & permitted access

Copilots · Enterprise search · Customer assistants · Tool-using agents