Threat modeling AI systems
MAESTRO for agentic systems and MITRE ATLAS techniques: mapping where an attack could get in before anything is built.
Training
Hands-on AI Security.
For security and engineering practitioners. Attack and defend real AI systems: assistants, retrieval, agents and tools, in an isolated testing range on synthetic data.
- For practitioners
- Attack and defend
- Live online
Who it's for
Built for this role.
- Security engineers, penetration testers and red teamers moving into AI.
- Developers and architects building with AI models, retrieval and agents.
- Security operations and assurance teams who will review AI systems.
Module 1 Module 2 How AI systems are built
Models, retrieval (RAG), tool calling and agents, and where the trust boundaries sit. Secure by design: controls planned in from the start, not bolted on.
Module 3 Prompt injection, direct and indirect
Attack an assistant and a retrieval system, then harden them. Mapped to OWASP LLM01.
Module 4 Data and model attacks
Data poisoning, model extraction, and supply chain risk in model files and third-party components.
Module 5 Agents and tools
Excessive agency and connector permissions. Zero-trust architecture for agents: never trust an agent's input or output by default, and verify every tool call.
Module 6 Defenses
Guardrails, least privilege, logging and monitoring for AI. Defense in depth: several layers, so one bypassed guardrail does not mean a breach.
Module 7 Red teaming and reporting
Scope a test, log every attempt, map results to OWASP and MITRE ATLAS, and score findings with OWASP AIVSS.
Format and outcomes
What participants leave with.
Dates on request.
Outcomes
- Threat model an AI system with MAESTRO and MITRE ATLAS.
- Run and document prompt injection, poisoning and agent attacks.
- Design layered defenses and prove them with a retest.
- Write findings a security team or auditor can act on.
Format
- Live online sessions for teams and individuals.
- Led by a senior, CISSP-certified lead.
- Grounded in OWASP, MITRE ATLAS and NIST guidance.
- Labs run in Agentiv Range, Agentiv's own testing range for AI systems, in isolated containers on synthetic data.
Before you start
Comfort with a command line and basic scripting helps.
Ready to get started?
Tell us who needs training and when.