You already know: IAM and least privilege (D17); the ML lifecycle from data to deployment (D11); data pipelines and dataset releases (D16); overfitting and generalization (D3); every input channel Atlas now accepts (D18 bridge).
Why this matters for Atlas. Classical security still applies to AI systems (patch servers, protect secrets, authenticate users), but ML adds new assets and new attack surfaces. Training data can be poisoned so the model learns a backdoor. Model files can carry code that runs when loaded. The model itself can leak its training data or be copied through its API. And LLMs read untrusted text as part of their input, which the next two lessons cover in depth. This lesson gives the method (threat modelling), the catalogue (OWASP's LLM risks and the ML-specific attacks), and working demonstrations of poisoning, unsafe deserialization and membership inference.
1. Threat modelling an ML system
The question. Where should a security review of Atlas even start?
Start from nothing. A threat model answers four questions (Shostack): what are we building, what can go wrong, what will we do about it, did we do a good job? Practically: draw the data flows, mark the trust boundaries (where data crosses from a less-trusted to a more-trusted zone), list the assets (things worth stealing or corrupting) and, for each boundary, ask what an attacker could do there. STRIDE (spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege) is a checklist for the "what can go wrong" step. For ML, add the assets classic models miss: training data, model weights, prompts and system instructions, embeddings and vector stores, evaluation sets, and the model's behaviour itself.