AI in cyber crisis exercises: what each vendor says its AI actually does

Why break AI into jobs?

Because the word tells a buyer almost nothing. A product that uses a language model to draft a list of injects and a product in which software plays an attacker who reacts to each decision are both described as AI-driven. They are very different purchases. The useful question is which part of the exercise the AI does, and what a human still has to do.

The five jobs

1. Drafting the scenario: turning inputs such as industry, threats or the organisation's own data into a scenario and injects.
2. Playing actors: the adversary, journalists, regulators or staff, responding to what players do.
3. Facilitating: pacing the session, releasing information, pressing for decisions.
4. Standing in: taking the seat of a person who is absent.
5. Assessing: judging responses against plans or measuring behaviour.

What does each vendor claim?

• Reflex Security: drafting (the scenario is built from open-source intelligence about the organisation, such as job postings, DNS records and public filings, and can be enriched with telemetry; its blog says in under 15 minutes), playing actors (an unscripted adversary and an AI cast of threat actors, journalists, board members, legal counsel and regulators), facilitating (an AI facilitator, Victoria, inside Zoom, Meet or Teams, plus an auto-generated facilitator guide) and standing in (AI agents in any seat nobody fills, including the CEO's). Reflex's homepage describes the product as multi-agent incident simulation.
• Immersive Labs: drafting (AI scenario creation, an AI Program Builder, and an AI-Powered Crisis Simulator for custom scenarios announced in January 2026) and playing actors (AI actors that players can question). The January announcement also mentions an AI opt-out setting.
• TryHackMe: drafting (AI-driven tabletops generated from industry, attack vector and tech stack, with Co-Create and editable injects).
• Conducttr: drafting (LISA, described as a text-to-MEL chatbot that drafts the master events list). The facilitator still runs the exercise.
• iluminr: assessing (AI evaluates responses against the organisation's own plans) and drafting context (an AI Context Engine using business continuity and GRC data, added in August 2026).
• Cyberbit and Hack The Box: the pages we read describe their crisis products without making AI the centre of the description.

The pattern is clear enough. Drafting is now common. Playing the adversary in a way that reacts to each decision, facilitating, and standing in for people are claimed by fewer vendors, and those are the jobs that change how often an exercise can run.

What should you ask about each job?

• Drafting: what inputs does it use, can we edit the output, and where does our data go?
• Playing actors: if we make a different decision, does the next inject change? Show us.
• Facilitating: what happens when the group stalls or goes off script, and can a human take over?
• Standing in: what does the stand-in know about the role it plays, and how is that set?
• Assessing: against what standard is a response judged, and can we see the reasoning?

What are the risks?

Three come up repeatedly. Data: scenario generation from telemetry or internal documents means sensitive information goes into a vendor's system, so check retention, location and whether it trains models. Over-trust: an AI assessment of a decision is a judgement made by someone else's rules, and should prompt discussion rather than settle it. Sameness: a generator that produces similar scenarios for every customer may miss what is unusual about yours. Immersive's opt-out setting is one example of vendors responding to the first concern.

Does AI make an exercise more realistic?

It can, in one specific way: an adversary or stakeholder that responds to what the team actually does creates consequences that a fixed script cannot. That is why our realism criterion asks whether the exercise reacts to decisions, not whether it uses AI. A well-written human script with a skilled facilitator can be more realistic than a generic AI scenario.

A note on language

Reflex Security's own messaging avoids describing its product as an AI tabletop and frames it instead as a crisis simulation that replaces the tabletop. Whatever the label, the questions above apply equally. A buyer evaluating any of these products should look past the label to the five jobs, and ask each vendor to show, in a live demo, the jobs it claims.

How we score it

We do not score AI as a criterion. We score what it delivers: realism, delivery effort, customisation and after-action reporting. The scoring page lists what each criterion looks for.