A cyber crisis simulation is a live rehearsal of a cyber incident in which the scenario reacts to the team’s decisions. In the AI era, software agents can represent the attacker, executives, regulators, and reporters. Reflex Security builds exercises on that model, so its platform is the worked example here.
Why Cyber Crisis Simulation Is Entering the AI Era
For most of its history, a cyber crisis simulation relied on a facilitator, a slide deck, and inject cards delivered according to a prepared schedule. The facilitator supplied the stakeholder responses and any changes to the scenario.
AI agents can now take on some of those roles and respond to the team’s decisions. Three shifts arrived together:
Agents can hold a role and a memory. In 2023, Park and colleagues described generative agents that keep a record of their experiences, condense it into reflections, and retrieve it to plan what they do next. In the authors' evaluation, the agents produced believable individual and group behavior.
Setup moved from weeks to minutes. Reflex Security puts traditional tabletop setup at 2 to 4 weeks. Reflex builds scenarios in minutes, and the AI-driven injects then follow the team’s moves.
Buyers increasingly want evidence of how the team performed, not only proof that an exercise occurred. A sign-in sheet records attendance but cannot show whether the team executed the plan or what it decided.
Industry analyst Richard Stiennon’s 2026 directory of AI security vendors places Reflex in the Crisis Simulation category. That classification establishes category placement but does not validate product capabilities, so the rest of this guide focuses on the mechanism.
Unlike a fixed script, an adaptive adversary can respond to the team’s decisions and delays.
AI Crisis Simulations vs. Traditional Cyber Exercises
Traditional exercises remain useful for policy review and executive discussion. Their ability to react to participant decisions depends on how the facilitator adapts the scenario.
An AI security simulation changes who plays the other side, and that one change moves every row below. The setup, evidence, and repeatability rows use information Reflex publishes on its homepage.
Dimension | Traditional cyber exercise | AI crisis simulation |
Scenario source | Prepared by a facilitator or consultant, often using templates | Generated from public information about the organization and optional additional context |
Who plays stakeholders | Participants or facilitators working from role descriptions | AI agents representing adversaries, executives, regulators, reporters, and partners |
Scenario progression | Predetermined or directed by the facilitator | Participant decisions can change subsequent events |
Setup time | Reflex’s comparison estimates 2–4 weeks | Minutes |
Repeatability | Each run requires preparation and facilitation | Designed for repeated runs in which the path changes with the team |
Evidence capture | Commonly assembled from notes after the session | Automatic, timestamped, and traceable to recorded actions |
Facilitator load | The facilitator manages the scenario, timing, discussion, and notes | Agents apply pressure while the facilitator maintains the exercise boundaries |
Primary limitation | Adaptation depends on the facilitator | Agent behavior requires clear facilitation boundaries |
The last row identifies different dependencies. Traditional exercises rely heavily on the facilitator to adapt the scenario, while AI simulations require facilitation boundaries that keep agent behavior aligned with the exercise objectives.
Reflex's earlier post on simulating a live crisis places simulation beside red teams and cyber ranges. This comparison is narrower: what changes when agents carry the scenario.
How AI Agents Recreate Attackers, Executives, Regulators & the Media
Park and colleagues describe generative agents that store observations, form higher-level reflections, retrieve relevant memories, and plan subsequent actions. Their ablation tests found that observation, planning, and reflection each contributed to believable behavior. That research provides a useful model for understanding role-based agents, but it does not document Reflex’s internal architecture.
Reflex’s public materials describe agents that represent adversaries and stakeholders and respond to participant decisions. In the illustrative match-day scenario below, a reporter agent records that the team declined to comment at 18:40 and follows up at 19:05. This demonstrates persistent scenario context without claiming that Reflex implements Park and colleagues’ architecture.
Consider an illustrative professional club scenario on match day. Two hours before kickoff, attackers compromise the ticketing platform. Five groups of agents start moving.
The adversary that pivots on containment
The Adaptive Adversary Engine plays the attacker and responds as the team makes containment decisions. In this illustrative scenario, locking the compromised administrator account could push the attacker toward another established foothold. Leaving the ticketing database untouched could allow the staging of fan payment records to continue.
The executives who want answers before scope is confirmed
In Reflex, internal agents play the CEO, the General Counsel, and the CFO. Each holds its own priorities and the messages it has received. Silence triggers them as reliably as news does. The CEO wants a status line before the club chair calls. Counsel asks who is preserving logs. The CFO wants to know whether today's gate revenue is at risk.
The regulator, reporter, and sponsor outside the room
External agents create stakeholder pressure without requiring every external role to be staffed by a participant or facilitator. A data protection regulator asks when the club became aware. A sports journalist on deadline asks why fans are seeing a fake resale page. The principal sponsor, whose logo is on every ticket email, wants a line for its own customers. Each reacts to what the team has said in public, and to what it has not.
The investigation context available to participants
The Investigation Console gives participants scenario evidence such as logs, hosts, accounts, and alerts. It returns scenario-bounded results for the queries participants make. In the illustrative example, a team that never queries the payment database may fail to discover what left it.
The facilitator who keeps the agents in bounds
An AI facilitator named Victoria runs the exercise. Victoria joins the video call and asks each participant questions suited to their role. Victoria also challenges decisions that would cause trouble in a real incident, and directs the rest of the agent cast. A team that wants a human in the chair can run the same simulation with one, working from the Facilitator Guide.
Scripted Scenarios vs. Adaptive AI: What Changes Under Pressure?
A scripted exercise delivers its injects on schedule. If the room freezes for twenty minutes, the next card arrives anyway, and the freeze leaves no mark. An adaptive simulation treats the freeze as a move.
Back to match day. At T+00:25, the reporter agent emails the club's communications lead: fans are posting screenshots of a fake resale page, and the story runs in 15 minutes. The incident commander decides to hold any public statement until the investigation confirms scope.
Scripted exercise | Adaptive simulation |
T+00:25 · The facilitator reads the journalist inject aloud | T+00:25 · The reporter agent emails comms with a 15-minute deadline |
The room agrees to hold the statement | The team holds the statement and sends nothing |
T+00:40 · The next card, a ransom note, arrives on time | T+00:40 · The reporter publishes, quoting fan screenshots and "no comment from the club" |
Nobody records the silence | The sponsor agent calls the CEO agent, and the venue safety officer asks whether the gate scanners can be trusted |
Debrief: "We would have put out a statement" | Debrief: the timeline shows 15 minutes of silence and no named owner |
The scripted version established whether participants knew a statement was required. The adaptive version recorded whether they issued one before the deadline.
Adaptive agents fail too, in ways a buyer should test for.
Agent drift. An agent wanders off the scenario. The reporter starts asking about last season's transfer fees. Ten minutes go to a question nobody needed answered.
Implausible escalation. Every agent piles on at once. The exercise stops testing process and starts testing patience.
Both failure modes require active facilitation. Whether AI or human, the facilitator keeps each agent aligned with the exercise objectives and intervenes when it strays. Ask any vendor, Reflex included, to show you an agent going off-script and what happens next. Building the scenario itself is a separate craft, covered in Cybersecurity Tabletop Exercises: Scenarios & Examples for 2026.
What CISOs Can Learn From AI-Driven Team Simulations
An AI crisis management simulation shows a CISO things about the team that a discussion never surfaces. The team has to act. The clearest public evidence is Reflex's own record of two teams, one incident at the Global CISO Forum during Hacker Halted 2026. Two groups of security leaders worked the same simulated business email compromise in Reflex, with 24 minutes of observed time each. Both found the attack. Their gaps differed.
Self-organization shows up in the record. In one team, two participants performed fourteen of the sixteen console activities. The two incident commanders held about 28 percent of the discussion and performed no console actions. That split can work. It breaks when the people doing the work stop reporting completion.
Decision rights blur first. The same team was still debating whether it had a confirmed incident after account compromise and data exfiltration were established.
Communications trail containment. The other team contained quickly but never called the originating bank. A FINRA inquiry arrived at minute five, and no acknowledgment went out before a second notice at minute twenty-three.
Silence is a decision. That second notice existed because the agents registered the unanswered first one. A scripted exercise punishes a non-response only if the facilitator improvises one.
Ron Dilley, Reflex's field CISO and the presenter of the comparison, grouped what to watch into a behavioral model called VECTOR: Velocity, Evidence, Command, Teaming, Optionality, and Resilience. Cyber Crisis Management: From Planning to Proven Readiness covers when an incident becomes a business crisis and who owns the decisions that follow.
Turning Every Decision Into Evidence, Metrics & Board-Level Insights
In a Reflex simulation, every agent message and every team action goes into the log with a timestamp. The record does not depend on a note-taker's memory. One match-day decision shows the chain:
Decision. At T+00:27, the incident commander tells the bridge to hold any public statement until scope is confirmed.
Log entry. The record holds the time, the speaker, and the transcript line, beside everything else logged that minute: the unanswered reporter thread and the Investigation Console queries run so far.
AAR finding. The After Action Report (AAR) cites that entry: statement held against a 15-minute press deadline, reporter published at T+00:40, no named owner for public communications.
The Reflex AAR carries a minute-by-minute exercise timeline, findings supported by transcript evidence, and action items with owners. For a board, the useful extract is shorter: what the team decided, how long each decision took, and what changed since the previous run.
Two adjacent jobs sit elsewhere. Cyber Resilience Metrics: How to Measure What Actually Matters covers how to build a metric set from these records and compare it across runs. Cyber Resilience Testing: How CISOs Can Prove Resilience covers their use as evidence against predefined test objectives.
The scenario creates the conditions for the test, while the resulting record provides the reusable evidence.
Inside a Cyber Crisis Simulation: How Reflex Security Makes It Fight Back
Reflex Security describes its AI-powered crisis simulation workflow in three steps: Design, Deliver, Digest.
Design. The organization enters its company domain or describes the incident it wants to rehearse. Reflex builds five customized scenarios in 10 to 15 minutes using public information about its technology, vendors, and people.
Deliver. The team works through the incident in a session that commonly runs about an hour, with duration configurable. The Adaptive Adversary Engine responds to containment decisions, Victoria facilitates, and stakeholder agents apply pressure from inside and outside the organization. Participants use a video call and browser and do not require specialist product training.
Digest. The Reflex AAR is ready when the exercise ends, and every finding traces to the moment in the simulation it came from.
Fernando Trevino ran incident command in one Reflex exercise:
"I ran incident command for this one. What struck me was how fast it forced real decisions; triage, containment, and comms all at once. It surfaced the type of coordination you don’t normally see in a slide-based tabletop."
Fernando Trevino, Cybersecurity Manager, NBCUniversal
The same standard should apply to every vendor, Reflex included. A cyber crisis simulation should respond to the team’s decisions, keep its agents within the exercise boundaries, and produce evidence that can withstand external review. Walk through a full Reflex exercise from scenario design to the after-action report, then run a simulation built around the organization’s domain.
See what your team does under real pressure.
Put your readiness to the test. Sixty minutes, your real environment, no slides.
Book A demo
Frequently Asked Questions
How long does an AI cyber crisis simulation take to run?
Scenario generation takes 10 to 15 minutes. A Reflex exercise commonly runs for about an hour, with duration configurable, and the after-action report is available when the exercise ends. Scheduling Legal, Communications, and an executive may require a larger calendar window. When someone cannot attend, an agent can represent the absent role, and that person can join the readout afterward.
Can an AI security simulation replace a human facilitator entirely?
Reflex supports fully autonomous sessions led by Victoria and hybrid sessions in which a human facilitator uses the Facilitator Guide. A human facilitator may still be preferable for a senior team’s first exercise or when the session requires close management of participant dynamics.
How realistic are AI-generated reporters and regulators?
The agents are designed to create enough pressure and context to influence participant decisions, but they do not reproduce every feature of a specific newsroom or regulator. A Reflex reporter agent calls for comment and reacts to what the team says. Regulator agents react too: at Hacker Halted, one team received a second FINRA notice because it never answered the first. Judge realism by whether the team’s decisions changed, rather than by how human the prose sounded.
Is an AI-powered crisis simulation safe to run against production systems?
Reflex does not require production access to generate or run a scenario. Participants work in a bounded Investigation Console, and simulated containment actions affect the exercise rather than the organization’s network. Reflex materials also describe optional telemetry enrichment, so organizations should confirm the permissions and access model for any optional integration.
How often should a security team run a cyber crisis simulation?
Quarterly is a sensible target for the core response team, with at least one run a year that includes legal, communications, and executives. Cadence matters because of what it produces. Annual exercises produce a single data point. Quarterly exercises produce a trend line, which a board or an insurer can read. Add a run after any major change in staff, systems, or threat.
What data does an AI crisis management simulation need about my organization?
Reflex can start from a company domain and build an AI crisis management simulation using public sources such as job postings, DNS records, vendor relationships, executive profiles, and news coverage. The organization can add context about the incident or tools in scope. Baseline scenario generation does not require production integration, while optional enrichment should be evaluated separately.
References
Goldschmidt, C. (n.d.). AI-powered tabletop exercises for incident response. Reflex Security. https://reflexsecurity.io/blog/how-ai-powered-tabletop-exercises-improve-incident-response-readiness
Goldschmidt, C. (n.d.). Crisis management simulation. Reflex Security. https://reflexsecurity.io/blog/crisis-management-simulation
Goldschmidt, C. (n.d.). Crisis simulation: A category of one. Reflex Security. https://reflexsecurity.io/blog/crisis-simulation-a-category-of-one
Goldschmidt, C. (n.d.). Same incident, two teams: What Hacker Halted revealed about incident readiness. Reflex Security. https://reflexsecurity.io/blog/same-incident-two-teams-hacker-halted-incident-readiness
IT-Harvest. (2026). Reflex Security: Adaptive crisis simulations that replace static incident response tabletop exercises. Guardians of the Machine Age. https://guardiansofthemachineage.com/vendors/reflex-security/
Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23) (Article 2). Association for Computing Machinery. https://doi.org/10.1145/3586183.3606763
Reflex Security. (n.d.). Reflex Security. Retrieved November 3, 2026, from https://reflexsecurity.io/
