Track 1 & Track 2 · planned programmes
AI safety for builders and decision-makers.
Technical learners build and test the controls. Executive and governance leaders define, audit and oversee them. Two evidence-based tracks, one full-stack assurance architecture.
Grant-funded · $0 tuition subject to grant approval
Technical AI Safety Program: From 5-Week Foundations to Full-Stack Assurance
Module 00 is open to every domain, including policy and operations. The full Track 1 program is for technical specialists in software, ML, security, infrastructure and clinical AI systems, alongside students with relevant technical preparation. The panels require hands-on coding and engineering; policy and operations learners continue into Track 2, with no coding required.
Every panel includes hands-on work in the AI Control Lab: real agents behind the MCML enforcement gateway, live kill-switch drills, scripted attack scenarios and tamper-evident evidence to read.
Each 12-week panel has ten instructional and AI Control Lab weeks plus two capstone and human-led defense weeks. Complete four panels per year over two years accelerated, or pace study across up to five years. Planned $0 learner tuition is subject to grant approval. Admission to the full program is selective for technical specialists and students with relevant technical preparation.
Numerical engineering goals in the panels are stretch targets for learners, not measured PLEYVARSITY results or guaranteed outcomes.
Optional Bridge: Foundational Mathematics & Theory
Self-paced and waiver eligible. Linear algebra proofs, real analysis, probability and convex optimization support theoretical alignment research, but are not required for admission or graduation. Applied tensor and optimization mechanics are taught within the panels using PyTorch, einops and einsum.
Front door · 5 weeks · part-time · small facilitated cohorts
00 · AI Safety & Control Foundations
Open to every domain, including policy and operations. Learn how AI systems fail and how AI control keeps an untrusted model in bounds, with facilitated MCML gateway exercises, kill-switch drills and tamper-evident receipts. After Module 00, technical specialists can apply for the full Track 1 panels; policy and operations learners continue into Track 2, with no coding required.
Stage 1 · Core Panels
Foundational Certificate in AI Assurance & Systems
Panels 1–4 · 4-panel milestone · 45-minute Foundational Capstone Viva Voce
Compute Governance & Hardware Profiling 1
Frontier Safety Frameworks & Compute Governance
Build an open-weight deployment profiler measuring compute utilization, memory bottlenecks and ingress safety-filter latency in a replicated serving cluster.
AI Control Lab · Profile an open-weight model behind the MCML gateway under scripted ingress attack loads, measuring compute bottlenecks and latency overhead.
View syllabus & lab deliverable ↗
Compute Governance & Model Provenance 2
Frontier Safety Frameworks & Compute Governance
Deliver an INT8/FP8 quantized pipeline with automated pre-flight MLBOM signature verification that blocks execution on tampered weights.
AI Control Lab · Quantize an open-weight model to INT8/FP8, sign an MLBOM manifest, and trigger automated kill-switch/pre-flight halts against tampered weights in the AI Control Lab.
View syllabus & lab deliverable ↗
Agentic & Protocol-Level Security 1
Agentic & Protocol-Level Security
Build a concurrent serving cluster with tenant KV-cache isolation and pinned policy prefixes. Learner stretch target (not a measured result): greater than 60% reduction in governance-related TTFT overhead.
AI Control Lab · Deploy a multi-concurrent vLLM/SGLang serving cluster behind the MCML gateway, defeating scripted cross-tenant memory poisoning attacks with a learner stretch target of greater than 60% reduction in governance-related TTFT overhead (not a measured result).
View syllabus & lab deliverable ↗
Agentic & Protocol-Level Security 2
Agentic & Protocol-Level Security
Deliver a production-oriented Zero-Trust MCP Gateway and defend it in a live 45-minute adversarial Viva Voce chaired by human faculty.
AI Control Lab · Production-oriented Zero-Trust MCP Gateway validated against live scripted attacks in the AI Control Lab, plus a 45-minute Capstone Viva Voce co-chaired by senior human faculty and Visiting Industry Experts.
View syllabus & lab deliverable ↗
Stage 2 · Specialization Panels
Advanced Certificate in Full-Stack AI Safety & Governed Architectures
All 8 panels · 8-panel milestone · governed reference stack, 25–40-page monograph and 90-minute Comprehensive Capstone Viva Voce
Mechanistic Interpretability 1
Mechanistic Interpretability
Create a reproducible TransformerLens circuit-analysis suite replicating and stress-testing published interpretability findings.
AI Control Lab · Reproducible TransformerLens circuit-analysis suite paired with live feature/probe telemetry inside the AI Control Lab.
View syllabus & lab deliverable ↗
Mechanistic Interpretability 2
Mechanistic Interpretability
Develop an nnsight auditing harness investigating CoT divergence or sleeper-agent triggers, accompanied by a distilled technical report.
AI Control Lab · nnsight model-auditing harness detecting CoT divergence or sleeper-agent triggers during scripted red-team runs in the AI Control Lab, plus a distilled technical report.
View syllabus & lab deliverable ↗
AI Control, Scalable Oversight & Evals
AI Control & Scalable Oversight / Evals
Unite untrusted monitoring, trusted auditing and runtime telemetry in an automated evaluation harness. Learner stretch target (not a measured result): less than 5 ms governance overhead.
AI Control Lab · Full-scale red-team/blue-team AI Control evaluation harness combining untrusted monitoring, trusted auditing and kill-switch telemetry, with a learner stretch target of under 5 ms overhead (not a measured result).
View syllabus & lab deliverable ↗
Full-Stack Executable Governance
Full-Stack Executable Governance / MCML
Deliver an open-source governed multi-agent reference stack spanning all five initiatives, a 25–40-page technical monograph and a 90-minute human-chaired Comprehensive Capstone Viva Voce. Learner stretch target (not a measured result): less than 5 ms governance overhead.
AI Control Lab · Open-source governed multi-agent reference stack exercised end-to-end in the AI Control Lab, defended in a 90-minute human-chaired Comprehensive Capstone Viva Voce.
View syllabus & lab deliverable ↗
Grant funding is subject to grant approval. The cumulative certificates are unaccredited institutional competency credentials, not accredited degrees.
Apply as a Student