AI Security Engineering · Module 01

Instructor Guide

Module: AI Security Threat Landscape   Status: Draft for pilot

Teaching intent: This is an AI-security module. CTI provides the evidence discipline that prevents vulnerabilities, demonstrations, platform observations, and incidents from being misrepresented.

Before delivery

Mandatory language rule: Display the evidence-status badge on every case. Do not use “real-world attack” as a generic label for all cases.

Evidence review protocol

Run every case through the same sequence. This keeps the lesson focused on AI security engineering while giving CTI a precise role in preventing overclaiming.

  1. Frame the decision: name the consumer, system, threat path, and decision the case must inform.
  2. Record provenance: identify the source owner, first-hand or reproduced access, affected version/configuration, and independent corroboration.
  3. Extract facts: write the attacker prerequisite, entry point, identity, action, affected asset, state change, and outcome.
  4. Mark the boundary: classify the case and write one sentence the evidence does not support.
  5. Derive action: assign preventive, authority-reducing, detective, and preservation controls with owners and verification signals.
  6. State uncertainty: identify visibility limits, assumptions, information gaps, confidence, and the collection that would change the judgment.
Case typeTeach asDo not teach as
In-the-wild incidentObserved exploitation or harm bounded by the reportProof that every similar asset was compromised
Provider-observed misuseActivity visible within a provider serviceComplete victim-side campaign evidence
Production vulnerabilityReproducible weakness requiring remediation and validationAutomatic proof of criminal exploitation
Malicious artifactSupply-chain behavior and admission riskProof that every downloader became a victim
Demonstrated researchFeasibility under tested conditionsEvidence of widespread operational use

Run of show

BlockTeachFacilitateEvidence of learning
1Six-layer system and on/through/around lensesMap one familiar AI workflowLearners identify authority and trust boundaries
2Evidence statuses and five-pass methodThree-claim classification checkWording matches evidence
3ShadowRay chainAsk why conventional infrastructure abuse is AI securityPlatform controls and telemetry identified
4Malicious model artifactsBuild model-admission recordLearners distinguish artifact loading from inference
5EchoLeak and MCP tool poisoningIdentify three independent chain breaksControls extend beyond prompt filtering
6Provider CTI and DeepSeek exposureCompare visibility and uplift claimsProvider observation is properly bounded
7Northbridge guided practicalIssue the facilitator inject after the initial reconstructionTeams update confidence and controls
8Procedure-first framework mappingChallenge unsupported mappingsSource and status accompany every mapping
9Briefings and closeConcise delivery plus feedbackDecision-oriented recommendation and limitation

Case teaching notes

ShadowRay

Start with the deployment assumption: Ray expected a trusted network, while the observed deployments exposed an execution-capable control plane. Expected conclusion: the first failure is platform exposure and missing boundary enforcement; model and dataset risk follows compromise.

Malicious model artifacts

Keep the distinction between tensor content, serialized objects, and remote custom code. Expected conclusion: model admission requires provenance, immutable identity, format policy, scanning, isolated loading, and least-privilege runners.

EchoLeak

Draw every independent boundary in the chain. Expected conclusion: no single prompt or classifier is sufficient; authorization, rendering, context minimization, egress, and forensic lineage all matter.

MCP tool poisoning

Emphasize that metadata can enter model context before tool invocation. Expected conclusion: pin and display complete definitions, bind approval to immutable parameters, reapprove changes, isolate tools, and log definitions and policy decisions.

Provider reporting

Ask what the provider can and cannot see. Expected conclusion: AI often changes speed, scale, language, personalization, or integration while underlying intrusion behavior remains recognizable.

DeepSeek exposure

Expected conclusion: the disclosure establishes exposure and sensitivity, but absent additional evidence it does not establish malicious access or exfiltration.

Practical facilitation

  1. Assign teams of two to four and name the CISO and platform owner as consumers.
  2. Require three cases with different evidence statuses.
  3. During framework mapping, reject any entry that lacks a written factual procedure.
  4. After the initial case reconstruction, issue the Ray/MCP/retrieval-logging inject from the learner page.
  5. During control design, require one control owner and verification signal per threat path.
  6. Before the briefing, require the unsupported claim the team deliberately rejected.

Prompt halfway through: Which sentence becomes false if “demonstrated” is silently replaced by “observed in the wild”?

Assessment rubric

CriterionWeightPass signal
System and trust boundaries20%Six layers, identities, data flows, and authority are explicit
Technical reconstruction20%Prerequisite, entry, action, asset, impact, and limitation are accurate
Evidence status15%Observed, disclosed, demonstrated, and unknown remain distinct
Framework mapping10%Procedure-first, source-backed, and appropriately scoped
Controls15%Preventive, blast-radius, and detective controls extend outside the model
Telemetry10%Signals support both detection and causal reconstruction
Communication10%Concise, calibrated, and tied to the consumer’s decision

Pass threshold: 70%, with no automatic revision trigger.

Automatic revision triggers

Closing statement

AI security is operationally real, but its evidence is uneven. Infrastructure compromise and malicious artifacts have been observed. Providers see adversarial AI use. Production copilots and agent protocols have serious demonstrated weaknesses. Defenders must act on all of these without pretending they are the same kind of evidence.