AI security leadership simulator

RedOps

Six months in the chair of an AI Security & Cyber Defense Lead at a financial firm adopting AI fast. Every decision is a tradeoff between security and the business.

Play the demo

The problem

Most organizations are adopting AI faster than their security programs can govern it. The hard part isn't the technology. It's the judgment calls: which use cases to approve, which controls to fund, and how to say no to an executive without becoming the department of no.

RedOps turns those calls into a turn-based campaign. You run intake, policy, the SOC queue, incidents, and executive meetings, and your decisions compound month over month.

What it does

Program scores that push back: Nine program scores, including security posture, shadow-AI risk, exec trust, budget, and regulatory exposure. Every decision moves several at once, and nothing is free.
A manager-level SOC queue: Escalated alerts and branching incidents whose outcomes depend on controls you put in place months earlier. Limited attention each month forces you to delegate.
Framework-mapped: Decisions map to OWASP LLM Top 10, NIST AI RMF, MITRE ATLAS, and Google SAIF, with detection-logic snippets and risk assessments throughout.
Graded outcome: Monthly reviews and a final report graded from Fired to RedOps Elite, exportable as Markdown or JSON.

Screens

RedOps campaign overview, month one of six

RedOps Policy Center with modular controls

AI features

AI-generated SOC escalation: suspected RAG poisoning of a policy wiki — Live mode: an AI-authored escalation built from current threat intel, with response options scored by the engine. Captured before live AI was paused on the public deployment for cost and abuse control.

Live threat briefing and an AI executive meeting with a skeptical CMO — Live intel briefing with cited sources, and an executive meeting where the AI stakeholder pushes back on your plan. Captured before live AI was paused on the public deployment for cost and abuse control.

How it’s built

App: Next.js 16 · React 19 · TypeScript
State: Zustand, deterministic scoring engine
AI: Anthropic API, server-side only
Validation: Zod schemas on every AI response
Data: Supabase (Postgres, auth, spend ledger)
Quality: Vitest test suite · installable PWA

A rule-based engine is the source of truth. AI only authors scenarios and proposes score impacts; code validates, clamps, and applies them. Outcomes stay fair and reproducible, and the whole game works with AI turned off.

Live mode researches current AI-security incidents through web search, then generates the month's use cases, alerts, and incidents from that intel with citations. Classic mode is a hand-crafted, fully offline campaign.

Security & AI decisions

AI is advisory, never authoritative: Model output passes through schema validation, vocabulary normalization, score clamping, and tradeoff checks before it can touch the game state. Anything that fails falls back to the built-in library.
Hard spend controls: A durable budget ledger caps total AI spend, and a per-client rate-limit gate throttles interactive calls. If the ledger is unreachable, AI fails closed instead of running uncapped.
A real kill switch: One environment flag turns every AI path off. That's how the public deployment runs today: visitors get the offline Classic campaign and the product still makes sense.
Keys never reach the browser: All model calls live behind a server-only provider module. Swapping providers means changing one file.

What’s next

Competency mapping: turn play signals into an evidence-backed profile of a leader's strengths across governance, risk, incident command, executive communication, and prioritization.
Cohorts: an instructor assigns the same drills and campaigns to a team and gets a completion and quality dashboard. This turns RedOps into a team training tool.

Source available on request.