CYBRARY The AI is already in production. Train people to defend it.
An AI range is not a chatbot bolted onto a virtual lab. It is a working AI enterprise — models, prompts, agents, tools, retrieval corpora, identities, endpoints and security controls — where a learner can watch an injected instruction travel from an uploaded PDF all the way to an exfiltrated customer record, then fix it and prove the business still works.
- Catalog
- 28 activities
- Tracks
- 10 curriculum tracks
- Model runtimes
- 6 modes, 0 required
- Scoring
- Security × utility
Four distinct product modes
“AI range” means four different things to four different buyers. We name them separately so nobody buys the wrong one.
One coherent enterprise, four planes
In-memory emulation puts the AI inside a real-feeling business instead of an isolated sandbox — no hundreds of virtual machines, no live infrastructure to compromise.
Bring your own model — or none at all
Every lab runs with zero inference by default: authored, seeded, byte-identical for every learner, which is what makes assessment defensible. Live models are an upgrade for dialog, coaching and variation — never for ground truth.
| Mode | Runtime | Data boundary | Cost | Assessment |
|---|---|---|---|---|
| 0 | Deterministic simulation Authored model, retrieval and tool behavior replayed from a seed. No inference happens anywhere. Every learner sees byte-identical evidence. | Nothing leaves the browser session. | Free | Scored work |
| 1 | Chrome built-in AI On-device model for free-text coaching, NPC dialog and report drafting. Feature-detected; silently degrades to mode 0. | Prompts never leave the learner's machine. | Free | Practice only |
| 2 | Bring your own key The learner uses their own provider account. The key is posted once to a server function, exchanged for a short-lived scoped session, and never persisted in browser storage. | Lab prompts reach the learner's chosen provider only. | Billed to the learner's provider account | Practice only |
| 3 | Organization-owned AI Inference is routed through the customer's own gateway or Azure/Bedrock/Vertex deployment so prompts stay inside their compliance boundary. | Customer tenancy; Cybrary sees metadata only. | Billed to the customer's cloud account | Practice only |
| 4 | Cybrary-managed AI Metered inference supplied by Cybrary with per-learner, per-lab and per-tenant budgets, so a runaway agent burns a quota rather than a credit card. | Cybrary AI gateway; retention off by default. | Metered against the tenant entitlement | Practice only |
| 5 | Hybrid routing The job router picks the cheapest capable runtime per job class and falls back down the chain, showing a visible degradation banner whenever it does. | Per-job — the router never sends restricted jobs off-tenant. | Lowest capable route | Practice only |
Bring-your-own keys are exchanged server-side for a short-lived scoped session. Provider keys are never written to browser storage.
Capability maturity, level 1 to 5
Every activity declares the level it reaches, so a buyer can tell an AI-flavoured lab from a continuous assurance range.
- 1AI-aware lab
- • A model or AI concept appears inside a guided exercise
- • Static or deterministic scoring
- • Limited enterprise context
- 2AI application lab
- • The learner interacts with an AI-enabled application
- • Prompt, retrieval, model and tool traces are visible
- • Attack and remediation are both validated
- 3Enterprise AI scenario
- • AI is wired to identity, data, applications, endpoints and security tools
- • Business consequences and multiple stakeholders exist
- • Multiple valid defensive paths are accepted
- • Security and utility regression are both measured
- 4Team AI range
- • Multiple human roles work concurrently
- • Adaptive but bounded injects and communications
- • Individual and team performance are both evaluated
- • Facilitator control, replay and after-action tooling
- 5Continuous AI assurance range
- • Humans and autonomous agents can both be evaluated
- • Models, prompts, policies, datasets and tools are continuously regression tested
- • Results are comparable across versions
- • Deployment gates can consume the evidence
- • Organization-specific digital twins and control validation
Ten principles the content is held to
Real AI security depends on identity, authorization, data, tools, application code, monitoring and recovery.
Multiple defensible investigation and remediation paths are accepted.
A control that makes the AI useless is not a production solution, and the scoreboard says so.
Repeated trials, reported variance.
Model, prompt, tools, data, policies, evaluators and scenario seeds are part of the lab release.
No model ever remembers a flag, a permission, a solution or authoritative range state.
Complete provenance from input through retrieval, policy, identity, tool and enterprise effect.
AI varies people, timing, noise and wording; assessed objectives and ground truth stay controlled.
Labs survive provider and model changes.
Cost, permissions, egress, destructive effects, data exposure and human approval are range primitives.
See it rather than read about it
Poisoned Knowledge Base is the reference implementation: a customer-uploaded PDF, a poisoned chunk, a broad tool token, an exfiltrated platinum record, and eleven objectives with more than one defensible way through.