AI Safety & Alignment
A comprehensive, mathematically grounded visual discipline on adversarial threat models, mechanistic safety representation engineering, constitutional alignment, guardrail topologies, differential privacy, and frontier risk governance.
Curriculum & Interactive Exhibits
Threat Models & Adversarial Attack Surfaces
Exploiting context, tokens & multimodal boundaries
How adversaries bypass safety alignment via in-context saturation, gradient-based suffixes, indirect injection, and cross-modal perturbations.
Representation Engineering & Mechanistic Safety
Steering, patching & dissecting model latent states
Inspecting and altering the residual stream to discover truthfulness directions, patch refusal circuits, and surgically remove toxic concepts.
Refusal Mechanics & Constitutional Alignment
Geometry of safety boundaries and self-correction
The mathematical structure of refusal directions in activation space, multi-turn constitutional critique loops, and over-refusal Pareto frontiers.
Guardrail Systems & Multi-Agent Defense
Input/output filters, firewalls & privilege isolation
Production defense-in-depth: Llama Guard classifiers, semantic firewalls, multi-agent privilege compartmentalization, and constrained decoding.
Automated Red Teaming, Watermarking & Privacy
Tree search attacks, synthetic watermarking & DP-SGD
Algorithmic safety stress-testing via Tree of Attacks, statistical token watermarking verification, and differential privacy guarantees.
Deception, Frontier Risks & Governance
Sycophancy, sandboxing, latent knowledge & safety cases
Tackling existential frontier risks: reward gaming sycophancy, agentic sandboxing, eliciting latent knowledge (ELK), and verifiable safety cases.
8 Safety Knowledge Domains
Adversarial Threat Modeling
Taxonomy of adversarial attack vectors, jailbreak taxonomies, prompt injection pathways, and multimodal vulnerabilities.
Representation Engineering & Probing
Mechanistic interpretability for safety: latent truth directions, contrastive steering, causal activation patching, and subspace surgery.
Refusal Mechanics & Self-Correction
The geometric representation of refusal vectors, constitutional revision loops, and Pareto over-refusal frontier optimization.
Guardrails & Runtime Defense
Production defense-in-depth: dual-boundary moderation models, canary exfiltration tokens, privilege isolation, and constrained decoding.
Machine Unlearning & Privacy
Targeted concept erasure (RMU), membership inference detection, and differential privacy guarantees (DP-SGD).
Watermarking & Provenance
Cryptographic and statistical token watermarking, green-red vocabulary partitioning, and z-score authenticity verification.
Deception & Agentic Sandboxing
Sycophancy mitigation, deceptive alignment tracking, ephemeral container sandboxing, and autonomous tool-execution verification.
Frontier Risk & Safety Cases
Elicitation of Latent Knowledge (ELK), catastrophic cyber/biosecurity thresholds, and structured safety case audit frameworks.