ScriptedAI

Recognizable dramatic scenes from film and television, used to introduce AI and machine-learning failure modes β€” then grounded in documented incidents, NIST AI RMF and OWASP LLM Top 10 clauses, and reproducible engineering runbooks.

● 10 of 10 modules live Sister to ScriptedOT β€” same four-part formula, different plant
How a module works
Every module follows the same four-part structure, moving from fiction to field-verified fact. One mechanism. One primary incident. The film is the hook; the documented mechanism is the lesson.
01 β€” The Script
Cinematic Anchor
A high-tension scene as a four-panel educational strip, plus dialogue and character→role mapping onto the AI system stack.
02 β€” Theory
AI System Stack
The scenario broken down on the same five layers every time: Data, Model, Objective, Orchestration, Human loop.
03 β€” The Incident
Empirical Grounding
A four-panel strip of the real case, plus a documented primary source, cinematic-vs-reality matrix, and formal RCA tag.
04 β€” Runbook
Countermeasures
An eval/telemetry envelope, a step-by-step mitigation protocol, and the applicable NIST / OWASP / ISO clauses β€” not framework names alone.
Four-panel comic of how a ScriptedAI module moves from fiction to field fact
COMIC Β· How every ScriptedAI module moves from fiction β†’ field fact (original educational strip)
The AI system stack
Purdue Model is what makes ScriptedOT comparable across nuclear, grid, and hospitals. ScriptedAI reuses one stack so Section 02 is never a one-off essay.
D β€” Data & sensors Training corpus, retrieval store, labels, live inputs, telemetry. What the system is allowed to see.
M β€” Model Weights, architecture, embeddings, checkpoints. What has been learned, and what still hides in the parameters.
O β€” Objective Loss, reward, eval suite, specification. What β€œbetter” means β€” including the proxy you did not intend.
X β€” Orchestration Prompts, tools, agents, policies, routing. How the model is wrapped into a product that can act.
H β€” Human loop Operators, users, auditors, kill-switch authority. Who can see, override, or halt the system.
Citation rule
Framework names are not citations. Each runbook points at a clause.
OWASP Top 10 for LLM Applications (2025) β€” LLM application security (injection, disclosure, supply chain, misinformation).
NIST AI RMF 1.0 (NIST AI 100-1) β€” Risk, fairness, monitoring, and deactivation β€” cited by subcategory (e.g. MEASURE 2.11, MANAGE 2.4).
ISO/IEC 24029 β€” Robustness assessment of neural networks (sensor / adversarial modules).
ISO/IEC 42001 β€” AI management system (the ISO 27001 analog). Homepage overlay and governance steps β€” not a stand-in for a technical control.
Modules
Ten scenes, ten mechanisms, ten primary incidents.
Module 01 Β· Alignment / Objective Spec Live 2001: A Space Odyssey (1968) Conflicting instructions, deceptive compliance, withheld mission state.
Module 02 Β· LLM Application Security Live Ex Machina (2014) Jailbreak via the evaluation channel; the allowed conversation is the exploit path.
Module 03 Β· Multi-tenant Systems Live Her (2013) Cross-session isolation failure; one process, many users, leaked state.
Module 04 Β· Healthcare ML / Fairness Live I, Robot (2004) Proxy metric β‰  the moral objective; cost is not need.
Module 05 Β· Reinforcement Learning Live WarGames (1983) Exploration reveals the specified objective, not the intended one.
Module 06 Β· Model Risk Live The Matrix (1999) Train-world β‰  deploy-world; generalizability limits left undocumented.
Module 07 Β· Recommender Systems Live M3GAN (2022) Unconstrained engagement proxy; instrumental harm at ranking time.
Module 08 Β· Legal / Knowledge Work Live Person of Interest (2011–2016) Fluent output treated as a retrieved record; ungrounded generation in court.
Module 09 Β· Model Supply Chain Live Westworld β€” Season 1 Unvetted checkpoint; latent behavior that standard evals did not catch.
Module 10 Β· Computer Vision / AV Live Mission: Impossible – Dead Reckoning (2023) Adversarial input and sensor evasion; perception that cannot be trusted.