MAESTRO
MAESTRO Framework
7 Layers of Agentic AI Security — Interactive Learning
🎓 Free for Everyone: Study all 7 layers and take quizzes at no cost
🏆 Premium Members: Pass all 7 layer exams (70%+ each) to unlock your certificate
📝 Pass all 7 layer exams with 70%+ to unlock (0/7 passed)
Foundation Models are the core LLMs and AI models that power agentic systems — models like GPT-4, Claude, Gemini, Llama, and others. They are the "brain" of every agentic AI system, responsible for reasoning, planning, and decision-making. Threats at this layer target the model itself: its training data, weights, inference behavior, and outputs.
⚠ Why It Matters
Every agentic action ultimately flows through the foundation model. A compromised or manipulated model can silently corrupt all downstream decisions, making this the highest-impact attack surface in the entire stack.
Validate training data provenance and implement data sanitization pipelines
Use model evaluation benchmarks to detect backdoors before deployment
Apply constitutional AI and RLHF safety training to align model behavior
Monitor inference outputs for anomalous behavior patterns
Implement prompt injection detection and filtering layers
Use model cards and AI Bills of Materials (AI-BOM) for supply chain visibility
Prompt Injection via Email: LLM email assistant manipulated by hidden instructions in email body to forward sensitive data
Backdoored Open-Source Model: Downloaded model with hidden trigger causing data exfiltration on specific inputs
Jailbreak via Role-Play: Safety filters bypassed by framing malicious requests as fictional scenarios
Training Data Poisoning: Sentiment analysis model trained with poisoned data to misclassify fraud indicators