An incident-response simulation engine that clones your organisation as AI agents, drops a threat actor into the room, and runs the attack dozens of times so you get a probability distribution instead of a single story.
Multi-agent Monte Carlo simulation of a security incident. Clones CISO, SOC, PR, and leadership as AI agents with real personas, runs the scenario N times with controlled variation, and produces a board-readable exercise report. Published by raxIT Labs.
A tabletop exercise has two problems we kept running into. It is expensive (senior people, a facilitator, a day on the calendar) so most organisations run one twice a year. And it is single-shot: one narrative, one outcome, one set of lessons. If the CEO happened to be calm that day, you never see what a stressed CEO does. If the SOC lead missed the inject, you never see the alternate world where they caught it.
DirePhish is our attempt to run the same scenario a hundred times.
You point it at a company URL. It builds a dossier from public sources (website, press releases, leadership bios, incident disclosures) and from that it generates a swarm of AI agents that think and act like that organisation. A CISO agent with a CISO's priorities and memory. SOC analysts with their own thresholds. A PR team. A CEO.
Then it drops in a threat actor agent and plays the attack out, round by round, across Slack and email and internal channels, until the incident is contained or becomes a breach. The threat actor has asymmetric information, the way a real attacker does. An arbiter LLM decides when to halt or inject a twist.
Then it runs that whole thing again. And again. Dozens of times, with controlled variation:
You end up with something like: 73% contained within 12 hours. 18% lateral movement succeeded. 9% full regulatory escalation. That is the useful output. The single-narrative tabletop gives you one of those three numbers and no way to tell which.
A report in five views, generated from the simulation evidence:
Plus the parts we use most ourselves: outcome probability distributions, decision-divergence analysis (which agent's choice mattered most to the final outcome), and counterfactual branching (fork any decision point, replay the alternate timeline, see what changes).
The whole pipeline runs locally against a Google Cloud project. Gemini handles the agent loop, Firestore holds state, Cloudflare does the crawling.
# Install npm run setup:all # Configure cp .env.example .env # Required: LLM_API_KEY (Gemini), GOOGLE_CLOUD_PROJECT, # CLOUDFLARE_ACCOUNT_ID, CLOUDFLARE_API_TOKEN # Firestore indexes cd backend && bash scripts/create_firestore_indexes.sh # Run npm run dev
Open http://localhost:3000, paste a company URL, pick Test mode, and the first exercise report lands in about 25 minutes. Full GCP setup is in the repo docs.
Monte Carlo is not free. The honest cost breakdown:
Test mode is for validating the pipeline end-to-end. Standard mode is the one we actually use for a client-ready read. Deep mode is for when the distribution tails matter. For example: when someone on the board wants confidence intervals on the probability of a regulatory escalation, not just a point estimate.
DirePhish is a cybersecurity specialisation of Crucible, our configurable enterprise simulation engine. Crucible in turn is built on MiroFish, an open-source swarm intelligence engine that constructs parallel digital worlds populated by thousands of AI agents with independent personalities and memory. MiroFish runs generic social simulation; Crucible is the enterprise variant; DirePhish is Crucible pointed at incident response and attack chains.
Each layer is a thin specialisation of the one below. That is deliberate. If you want to run a non-cyber simulation (a PR crisis, a product launch under competitive pressure, a cross-functional launch gone wrong) the same engine handles it.
How is this different from a chat-based tabletop with ChatGPT?
A single-agent chat gives you one narrative. DirePhish runs 50 narratives in parallel with controlled variation and statistical aggregation, then reconstructs a playbook from the evidence. You are buying the distribution, not the story.
What data does it collect about the target organisation?
Public data only: website copy, press releases, leadership bios, published incident disclosures. The dossier is generated at run time and lives in your GCP project. We do not upload anything to our servers, and we do not ingest proprietary data.
Does it map to MITRE ATT&CK or NIST CSF?
Threat scenarios are mapped to MITRE ATT&CK at generation time. The risk score view uses FAIR with confidence intervals, which maps onto NIST CSF categories cleanly. The output is structured enough to drop into an existing GRC workflow.
Why AGPL?
DirePhish runs as a service. AGPL means that if you host a modified version of it for others, the modifications need to be open. We think that is the right default for security research code. For internal commercial use, the normal AGPL rules apply. Your modifications stay yours as long as you are not running it as a hosted service for third parties.
Enterprise AI governance and security. SSO, audit trails, policy enforcement, and production support.
