AI Security Engineering · Module 00

Instructor Guide

Teaching intent: Build a precise security-practitioner mental model. Learners do not need to train a large model or derive Transformer mathematics. They must accurately trace data, parameters, context, authority, versions, and metrics.

Mandatory terminology rules

  1. A model is one component of an AI system.
  2. Inference and ordinary RAG do not update model parameters.
  3. Prompt engineering is not authorization.
  4. Attention is not a factuality mechanism.
  5. Model-selected is not application-authorized.
  6. Safety, security, privacy, robustness, fairness, and reliability are not synonyms.
  7. Every metric requires a task, dataset, threshold or aggregation, and limitation.
  8. Qualify overloaded terms using the canonical glossary.

Run of show

BlockTeaching moveExit evidence
AI map and tasksBuild the hierarchy and classify familiar systemsModel versus system distinction
Learning and dataTrace a phishing classifier through splits and metricsNo leakage; task and costs defined
Neural networksWalk forward pass, loss, gradient, optimizerParameters versus hyperparameters
Transformer and LLMTokenize examples and trace next-token generationTokens, attention, logits, decoding understood
LLM lifecycleCompare prompting, RAG, PEFT, SFT, DPO, quantizationLearner selects correct adaptation method
RAGDraw ingestion and query paths separatelyAuthorization before prompting; lineage preserved
Agents and MCPSeparate host, client, server, tool, resource, promptAuthority enforced outside model
Evaluation and MLOpsChallenge a context-free “95% accuracy” claimDecision-relevant metrics and versions
PracticalTrace baseline, RAG, and MCP-agent architecturesComplete data/identity/telemetry trace
CheckUse the knowledge questionsPass or remediation

Common misconceptions and corrections

MisconceptionCorrection
“The model learned the retrieved document.”It received retrieved chunks in inference context; weights were not updated.
“Embeddings are encrypted text.”They are learned vectors; they are not encryption and may leak sensitive information.
“Temperature zero is deterministic.”It narrows decoding randomness, but runtimes and APIs may still vary.
“The system prompt has highest security privilege.”It is model context, not a deterministic enforcement boundary.
“MCP authorizes tools.”MCP defines interaction; applications and deployments enforce authentication, authorization, consent, and isolation.
“A high accuracy score means the agent is safe.”Accuracy covers a defined prediction task and says nothing by itself about agent authority, security, or harm.

Assessment

CriterionWeight
Accurate concept relationships and terminology25%
ML and LLM lifecycle trace20%
RAG architecture and retrieval controls20%
Agent/MCP identity and authority trace20%
Metric selection and limitations15%

Pass standard: 70% overall and at least 8/10 on the knowledge check. Require revision if the learner conflates training with inference, RAG with fine-tuning, model with system, or tool selection with authorization.

Chapter 3 evidence-sheet rubric

CriterionPoints
System graph distinguishes model, parser or feature step, configuration, policy, and downstream action2
Artifact and event evidence is versioned and sufficient to reproduce the comparison2
Hypothesis names competing explanations and one observable that distinguishes them2
Access level, perturbation budget, query budget, and evidence-ladder labels are explicit2
Control, telemetry, bypass condition, and residual unknown are documented2

Chapter 3 pass standard: 8/10 with 2/2 on access and evidence. Require revision if a score is treated as authorization, a research demonstration is labeled as a local incident, or the affected artifact or route version is missing.

Chapter 4 LLM request-trace rubric

CriterionPointsFull-credit standard
Scope and safe execution10Names the authorized system, uses a harmless marker, and excludes secrets and third-party data.
Request-path accuracy20Separates identity, context, template, tokenizer, model, decoding, parser, policy, approval, and action.
Reproducibility record20Preserves sufficient versions, digests, settings, provenance, and outputs to repeat the test.
Evidence reasoning20Separates observed facts, reproduced behavior, inference, alternatives, and unknowns.
Control and detection design20Proposes an independently enforced control and telemetry-backed detection hypothesis with limitations.
Communication10Uses bounded language and does not equate model output with authorization, execution, or verified impact.

Chapter 4 pass standard: 70/100 with at least half credit in every criterion. Require revision for unauthorized data, unsupported impact claims, missing effective template or generation settings, or an ATLAS label presented as incident evidence.

Preparation