Testing & Attack Simulation
We test whether your AI systems are resilient to attacks and compliant with regulations
We test LLM-based applications and AI systems for prompt injection, jailbreaks, data leakage and model extraction. We combine offensive testing with AI Act compliance assessment and GDPR data processing requirements. We work from the OWASP Top 10 for LLM Applications 2026 and, for AI agents, the OWASP Top 10 for Agentic Applications, using tools such as PyRIT, Giskard and promptfoo.
- 5
- 8+
Testing & Attack Simulation
- Security testing of chatbots and LLM-based applications
- Prompt injection testing
- Data leakage risk assessment
Chatbots and AI assistants reach customers and staff faster than organisations can assess the risk. A language model connected to company data and systems opens new attack paths: prompt injection, leaking another user's data, disclosing the system prompt, an agent that does more than it should. The AI Act sorts systems into four risk categories with documentation and transparency requirements, and the GDPR limits how models may process personal data. We test an AI system before it reaches users and combine offensive testing with a compliance assessment. At a financial institution preparing a customer chatbot, the pre-launch test found prompt injection paths to other customers' data; the team put the safeguards in place before go-live.
Scope of the service
Security testing of chatbots and LLM-based applications - aligned with OWASP Top 10 for LLM Applications 2026
Prompt injection testing - direct (manipulating the prompt itself) and indirect (injection via external data sources, e.g. documents in RAG pipelines)
Data leakage risk assessment - training data, user data, the system prompt and other hidden context (LLM08:2026 Hidden Context Exposure)
Data and Model Poisoning (LLM05:2026) - assessment of RAG sources, the fine-tuning data pipeline and model provenance for poisoning
Improper Output Handling (LLM10:2026) - validation and sanitization of model outputs before passing to downstream systems
Excessive Agency (LLM03:2026) and agent risks per the OWASP Top 10 for Agentic Applications (ASI01-ASI10) - AI agent permissions, tool chains, memory and trust boundaries
Unbounded Consumption (LLM06:2026) - resilience testing against excessive inference resource consumption (denial of service, denial of wallet) and model extraction through the API
AI Act compliance audit - risk classification (unacceptable, high, limited, minimal), documentation, transparency requirements
Personal data processing assessment for AI systems (GDPR)
Bias and fairness testing - identifying biases in model outputs
How we work, step by step
Scoping
We identify AI systems in the organization, their roles, input data, and integrations. We determine which systems fall under AI Act regulation and in which risk category. We agree on the testing scope.
Threat modelling
We analyse the AI system architecture, tool chains, data sources, and potential attack vectors. We map risks to the OWASP Top 10 for LLM Applications 2026, including prompt injection (LLM01), sensitive information disclosure (LLM02) and excessive agency (LLM03), and for agents also to the OWASP Top 10 for Agentic Applications.
Automated and manual testing
We conduct offensive tests using tools such as PyRIT (Microsoft), Giskard, and promptfoo. We test for prompt injection, jailbreak, data leakage, model extraction, and improper output handling. We combine automation with manual exploration of scenarios specific to the client's business logic.
Report
We deliver a report with prioritised findings, risk assessment, mapping to OWASP LLM Top 10, and actionable remediation recommendations. For high-risk systems we provide input to the technical documentation and conformity assessment.
Remediation support
We help implement defences - guardrails, input/output filtering, agent permission restrictions, and anomaly monitoring for model queries.
AI test or standard penetration test: which one you need
AI Security Testing
Choose an AI security test when the application uses a language model: a chatbot, an assistant, a RAG system or an agent with access to tools. We test prompt injection, data leakage and agent permissions against the OWASP Top 10 for LLM Applications.
Choose a penetration test for the application, API and infrastructure themselves. For a chatbot with access to customer data it is worth combining both, because classic application vulnerabilities remain alongside the model risks.
Choose a risk and compliance analysis when the organisation is still deciding on an AI service, such as Microsoft 365 Copilot, and needs an assessment against DORA, the AI Act and cloud requirements.
Who we serve
We've worked with national energy grid operators, systemically important banks, industrial automation manufacturers, renewable energy operators, and US DoD contractors.
Not sure this service fits your situation? Describe your environment and deadline in a few lines and we will propose a scope. We reply within one business day.
Frequently asked questions
What AI systems do you test?
We test chatbots, AI assistants, autonomous agents, RAG systems (Retrieval-Augmented Generation), applications built on OpenAI and Anthropic APIs, open-source models, and proprietary solutions. Any system that processes user data or makes decisions requires security testing.
What is the OWASP Top 10 for LLM Applications?
It is a list of the 10 most critical security risks for applications built on large language models, published by the OWASP GenAI Security Project; the current edition is 2026, released in August 2026. It covers prompt injection (LLM01), sensitive information disclosure (LLM02), excessive agency (LLM03), supply chain (LLM04), data and model poisoning (LLM05), unbounded consumption (LLM06), hidden context exposure such as system prompt leakage (LLM08), vector and embedding weaknesses in RAG systems (LLM09) and improper output handling (LLM10), among others. We base our testing on this list and, for AI agents, add the OWASP Top 10 for Agentic Applications.
Does the AI Act apply to my organization?
The AI Act applies to providers and deployers of AI systems on the EU market. The regulation defines four risk categories: unacceptable (e.g. social scoring, subliminal manipulation), high (e.g. recruitment systems, credit scoring, critical infrastructure), limited (chatbots - must inform users they are interacting with AI), and minimal (e.g. spam filters). We help assess which category your systems fall into and what obligations follow.
What tools do you use for testing?
We use PyRIT (Python Risk Identification Tool, Microsoft's open-source framework for red teaming generative AI), Giskard (an open-source library for testing model quality, bias and security), promptfoo (a prompt testing tool), and custom test scenarios. We combine automated tools with manual exploration, because many vulnerabilities - such as indirect prompt injection through documents in RAG - require a creative, context-aware approach.
How long does an AI security test take?
A typical project takes 2 to 4 weeks, depending on system complexity and the number of integrations. A simple chatbot without external integrations takes about 2 weeks. A RAG system with multiple data sources and autonomous agents requires 3 to 4 weeks.
What does the price of AI security testing depend on?
A project for a simple chatbot without integrations usually takes 2 weeks, and one for a RAG system with many data sources and agents 3-4 weeks; the number of working days follows from that. We multiply the days by the daily rate and break them down by stage, from threat modelling to remediation support. We agree the scope on a free call.
Related sectors, services and insights
Sectors
Services
- Penetration TestingWe show what an attacker could achieve through the flaws, and retest after your fixes.
- Red TeamingWe test whether your defence team detects and stops a multi-stage attack on your organisation.
- Security ArchitectA senior expert reviews your projects and change requests within a monthly pool of hours.
We'll discuss scope, methodology, and timeline.
Free consultation, no strings attached.








