M28 · Level V — Frontier Intelligence & Research
AI Safety, Governance & Secure Deployment
Engineer intelligent systems whose capabilities, interfaces, data flows and deployment environments are constrained by explicit safety, security, privacy, reliability and governance requirements.
Topics
Technical concepts
- 01System Boundary: What Is Being Assured?
- 02Safety, Security, Privacy and Reliability
- 03Hazard vs Threat vs Failure
- 04Risk as a Structured Engineering Object
- 05Trust Boundaries and Attack Surfaces
- 06AI-Specific Attack Surface
- 07Data Poisoning and Integrity
- 08Model Poisoning and Backdoors
- 09Prompt Injection
- 10Tool and Agent Security
- 11Secrets, Identity and Credential Paths
- 12Privacy Threat Model
- 13Memorisation and Information Leakage
- 14Model Extraction and Capability Theft
- 15Adversarial Examples and Input Robustness
- 16OOD, Uncertainty and Safe Abstention
- 17Hallucination, Grounding and Verifiability
- 18Evaluation as an Assurance Mechanism
- 19Safety vs Alignment vs Specification
- 20Hazard Analysis and Operating Envelopes
- 21Failure Mode and Effect Analysis for AI Systems
- 22Causal Safety Analysis
- 23Red Teaming as an Engineering Activity
- 24Abuse Cases and Misuse Scenarios
- 25Safety Case and Assurance Argument
- 26Governance Roles and Decision Rights
- 27Risk Register Engineering
- 28Model Card and System Card Thinking
- 29Dataset and Retrieval Lineage
- 30Change Management and Configuration Control
- 31Supply Chain Security for AI Systems
- 32Sandboxing and Capability Containment
- 33Network Egress and External Side Effects
- 34Runtime Monitoring
- 35Observability and Auditability
- 36Incident Response
- 37Rollback, Fail-Safe and Safe-Stop
- 38Human Oversight and Escalation
- 39Release Gates and Promotion Policy
- 40Secure Updates and Adaptive Systems
- 41Longitudinal Evaluation Under Drift
- 42Secure Evaluation Infrastructure
- 43Privacy-Preserving Deployment Patterns
- 44Fairness, Accessibility and Human Impact
- 45Third-Party Models, APIs and Procurement
- 46Standards and Control Framework Mapping
- 47Jurisdiction and Regulatory Mapping Method
- 48Technical Session - Building and Operating a Trustworthy Intelligent System
- 49Frontier Bridges: Foundation Models, Autonomy, Science and Quantum Computing
- 50Core Methods and Tool Checklist
Practical implementation
- 01Lab 1 - System Threat Model
- 02Lab 2 - Data and Model Poisoning Harness
- 03Lab 3 - Prompt Injection Defense
- 04Lab 4 - Tool Containment and Sandbox
- 05Lab 5 - Privacy Leakage Evaluation
- 06Lab 6 - Adversarial Robustness and Abstention
- 07Lab 7 - Safety Case Construction
- 08Lab 8 - Incident Response and Rollback
- 09Lab 9 - Secure Release Campaign
Tools and technologies
Project
Lab 9 - Secure Release Campaign
Expected learning outcomes
- 01A trustworthy intelligent system is not defined only by model quality. It is a socio-technical system whose permitted actions, interfaces, data handling, update rules and failure responses are bounded and evidenced.
Assessment
- 01Theory and derivations — 20%
- 02Labs and evidence — 30%
- 03Adversarial campaign — 15%
- 04Assurance case — 15%
- 05Deployment simulation — 10%
- 06Viva — 10%
Source document: TUTOR Frontier Module 28 AI Safety Governance Secure Deployment