A focused security review of one LLM application, RAG system, or AI agent: threat model, control review, and authorized adversarial testing.
AI systems introduce attack paths traditional application security testing does not cover: prompt injection through user input and retrieved content, sensitive data disclosure, excessive agent permissions, insecure output handling, and model and plugin supply chain risk.
The assessment examines how your system actually handles untrusted input, what it can reach, and what it is allowed to do — then validates the highest-risk paths with testing performed only under written rules of engagement.
Define the system, environments, and test boundaries; sign rules of engagement.
Map architecture, data flows, and trust boundaries; build the threat model.
Assess controls against the threat model and run authorized adversarial tests.
Severity-rated findings, remediation walkthrough with engineering, and retest.
It includes targeted adversarial testing, but the emphasis is design and control weaknesses specific to AI. It complements, rather than replaces, your standard application penetration test.
Only with explicit written authorization. The default is a non-production environment that mirrors production configuration.
Your integration: prompts, retrieval, tools, permissions, data handling, and logging. The provider's own infrastructure is covered through vendor due diligence, not testing.