What does LLM penetration testing cover?
Direct and indirect prompt injection, jailbreaks and guardrail bypass, data leakage through training data, embeddings and context, and abuse of agents and tool-calling inside your application.
Penetration Testing / AI & LLM
AI & LLM Penetration Testing
Security testing for AI systems: prompt injection, model abuse and data-leak paths.
What we test
Scope of testing.
- Direct and indirect prompt injection across assistants and RAG pipelines
- Jailbreaks, guardrail bypass and model abuse
- Data-leak paths: training data, embeddings and context exposure
- Agent and tool-calling abuse inside applications
Process
Mapped to the five phases.
The same discipline applies across every discipline we test.
- 01
Scoping
Define assets, rules of engagement, timelines and testing type (black / grey / white box).
- 02
Discovery
Reconnaissance, enumeration, threat modelling and attack surface mapping.
- 03
Testing
Vulnerability identification and controlled exploitation to confirm impact.
- 04
Reporting
Findings with CVSS scoring, evidence, business impact and remediation steps.
- 05
Re-test
Verification of fixes and issuance of a closure report.
FAQ
AI & LLM: common questions.
Straight answers on scope, method and deliverables.
Do you test RAG pipelines and chatbots?
Yes. Retrieved content is a classic injection path — we test how your pipeline handles hostile documents and how far an attacker can push a conversation.
Which tools do you use?
Open-source harnesses including Garak, PyRIT and Promptfoo, combined with custom test cases written for your application's specific tools, data flows and guardrails.
Scope this assessment.
Tell us the assets and constraints. We will come back with a tailored proposal.