1. The pitch deck
Twelve slides. Read each one out loud once before the meeting to lock the cadence. Open the corresponding screen on the second monitor when you say the slide.
SLIDE 01 · COVER
Mythal — the autonomous vulnerability remediation fabric for critical infrastructure
Twelve coordinated AI agents that close the loop from CVE to verified fix — without ever endangering an OT system.
Madhu Uppalapati · AI Architect · Next-Era LLC · For CSX
Open with: "What you're going to see in the next 25 minutes is a complete autonomous vulnerability remediation platform — running live, against a simulated Class I rail estate with 4,300 assets. Built from the assumption that AI-discovered CVEs are now the operating condition, not the exception."
SLIDE 02 · THE MYTHOS PROBLEM
Vulnerability discovery has gone machine-speed. Remediation has not.
163
CVEs Microsoft patched in April 2026 alone
~28%
of CVEs now AI-assisted (Mythos & peers)
< 24 hrs
from disclosure to active exploitation (avg, 2025)
48 days
average MTTR in critical infrastructure (Ponemon 2025)
The gap between when a vulnerability becomes weaponized and when an operator can patch it is now measured in hours on one side and weeks on the other. That delta is the largest unhedged risk on a Class I rail CISO's balance sheet — and it cannot be closed by hiring more humans.
Say: "If you remember nothing else, remember this — vulnerability discovery has become a machine speed activity, and remediation has not. The asymmetry is the entire investment thesis."
SLIDE 03 · WHY CRITICAL INFRASTRUCTURE IS WORSE
Rail, pipeline, power, water have it worse than anyone else
- Constrained maintenance windows. You cannot patch a wayside interface unit during a train movement. Most OT systems have one quarterly window of opportunity.
- Vendor-locked firmware. Siemens RTUs, Wabtec PTC boxes, Hitachi locomotive controllers — patching is vendor-coordinated, not push-button.
- Regulatory exposure. TSA SD 1580-21-01 requires documented timely patching of Critical Cyber Systems, with auditor-ready evidence.
- Asymmetric blast radius. A misfire on a yard SCADA system stops freight. A misfire on a server is contained.
Say: "This is why a generic IT vulnerability tool fails on a railroad — and why CSX has been telling vendors what you need is a Middleware AI Agent. We agree on the workflow. We disagree that one middleware agent is enough."
SLIDE 04 · WHAT CSX ASKED FOR
Your stated workflow — and the seven gaps we filled in
| What CSX asked for | What you also need (Mythal covers) |
|---|
| Receive trigger for critical CVE | Multi-scanner dedup across Qualys, Tenable, Wiz, Claroty, Defender |
| Assess threat level | KEV / EPSS / ransomware-actor enrichment + business-impact join to CMDB |
| Download appropriate patch | Vendor-aware resolver (MS bulletin, Cisco SA, Siemens SSA) with reliability score |
| Update system | OT Safety Officer veto · dual approval · executor through Tanium/SCCM/Ansible/Panorama · verifier rescan · auto-rollback |
| — | Signed reasoning trace auditors can read |
| — | Compliance evidence mapped to TSA SD 1580-21-01, NIST CSF 2.0, NIST 800-82r3, IEC 62443 |
| — | Mythos-aware — built assuming the patch firehose is the operating condition |
SLIDE 05 · THE MYTHAL FABRIC
Twelve specialist agents · one signed message bus · one policy gate
★ Supervisor
Orchestrator. Manages the per-finding state machine. Runs on Claude Opus.
Scanner Liaison
Normalizes findings from Qualys, Tenable, Rapid7, Wiz, Defender, Claroty, Nozomi, Dragos.
Threat Intel Aggregator
Enriches with NVD, KEV, EUVD, vendor PSIRTs, GHSA, ICS-CERT.
Patch Hunter
Resolves the vendor fix. MS bulletins, Cisco SAs, Siemens SSAs, vendor archives.
Impact Analyst
Joins to CMDB → business criticality, network exposure, blast radius.
Change Risk
Scores against historical change-failure rates · recommends window.
★ OT Safety Officer
Veto rights on OT and CCS assets. Recommends compensating controls. Runs on Claude Opus.
Remediation Planner
Produces ordered runbook + machine-executable workflow + rollback.
Executor
Applies via Tanium / SCCM / Ansible / Catalyst Center / Panorama / OT-native.
Verifier
Re-scan · health check · exploit re-test · auto-rollback on failure.
Compliance Reporter
TSA SD 1580 · NIST CSF 2.0 · NIST 800-82r3 · IEC 62443 evidence packages.
★ Inventory Insights
Asset graph maintenance, dependency discovery, CMDB delta detection.
Specialists don't share memory. They communicate over a typed, signed message bus. Every decision lands in the reasoning-trace ledger. Every side-effecting tool call passes through a deterministic policy gate first.
SLIDE 06 · THE OT SAFETY OFFICER
The single most important agent for selling into rail
Holds veto rights on every action targeting an asset tagged Critical Cyber System or sitting in an OT zone. Default policy: no direct firmware patching during operations. Instead:
- Compensating controls now. Tightens industrial-firewall ACLs · pushes IPS signature · puts the device under monitored isolation.
- Firmware patch scheduled. Pinned to the next planned maintenance window, with dual approval (security + OT Operations) required.
- Mapped to standards. Every veto records the NIST 800-82r3 + IEC 62443 zone-conduit rationale.
Say: "On a railroad, the difference between buying Mythal and buying a generic vuln-management tool is one agent — this one. Without an OT Safety Officer, the tool will eventually patch the wrong device at the wrong time, and a train will stop."
SLIDE 07 · THE CLOSED LOOP IN 11 STAGES
From CVE landing in a scanner to a signed compliance evidence unit
Each stage owned by one agent. Each transition signed. Each side effect policy-gated.
/how-it-works ↗ shows the deep dive in the console.
SLIDE 08 · WHAT WE READ FROM YOUR ESTATE
Scanner-agnostic. Patch-tool-agnostic.
| Category | Sources we plug in |
|---|
| IT scanners | Qualys VMDR · Tenable.io · Rapid7 InsightVM · Wiz · Microsoft Defender VM |
| OT scanners | Claroty xDome · Nozomi Guardian · Dragos Platform |
| Threat intel | NVD · CISA KEV (live) · EUVD · GHSA · ICS-CERT · MS / Cisco / Siemens / Wabtec / Hitachi PSIRTs |
| CMDB / inventory | ServiceNow · Device42 · BMC Helix · Mythal-native asset graph |
| Patch tools | Tanium · SCCM · Intune · Ansible · BigFix · Catalyst Center · Panorama · OT-native (RUGGEDCOM, Wabtec) |
| Identity | Okta · Auth0 · Entra · Keycloak · SAML SSO |
| Ticketing | ServiceNow ITSM · Jira |
Click any of the 17 cards on the Integrations page — modal shows authentication mode, polling cadence, the exact fields we read, and how each field maps to Mythal's canonical schema (e.g. QID + CVE list → vulnerability.cve).
SLIDE 09 · LIVE EXECUTION
You see the patch land — step by step
When a plan is approved, the Executor agent runs each step through the appropriate patch tool with realistic timing. The plan-detail page streams the timeline in real time: tool name, agent ID, started/completed timestamps, key-value result payload (Tanium action_id, SCCM deployment_id, Cisco DNAC task_id), and the Verifier rescan verdict per step. A 4-step plan completes in 8–12 seconds with visible progression.
If a step fails (18% failure rate on OT for realism), Mythal rolls back automatically and the plan transitions to ROLLED_BACK.
SLIDE 10 · COMPLIANCE EVIDENCE
Auditor-ready PDFs in under 60 seconds
- TSA SD 1580-21-01 — primary framework for Class I rail. Control-by-control evidence with execution + approval signatures.
- NIST CSF 2.0 — Identify / Protect / Detect / Respond / Recover, mapped to plan trace IDs.
- NIST 800-82r3 — ICS-specific. Zone-conduit evidence from OT Safety Officer decisions.
- IEC 62443 — IACS security. Compensating-control evidence with HMAC signatures.
One click on the Compliance page produces the PDF. The reasoning trace is the audit log.
SLIDE 11 · ROI
What CSX would save in year one
| Lever | Conservative | Aggressive |
|---|
| Vuln analyst FTE deflection | 3 FTE × $185K | 6 FTE × $185K |
| Mean time to remediate (days → hours) | 14× faster | 40× faster |
| Cyber insurance premium reduction | 5% | 12% |
| Audit prep effort | ~$400K / yr saved | ~$900K / yr saved |
| Avoided incident (per-incident) | — | $4–18M typical for rail |
ACV: $750K – $3M depending on tenant size, integration count, on-prem vs SaaS.
SLIDE 12 · ASK
What we need from CSX
- A 14-day technical proof in your lab environment — one Qualys instance, one OT scanner, one zone of 200–500 assets. Mythal stands up an on-prem appliance, connects to your sources, and runs the closed loop on real findings.
- A named OT Operations approver and Security approver for the dual-approval flow. Your existing role definitions are fine.
- Read-only access to one CMDB instance so we can demonstrate the business-impact join.
- One auditor (TSA or internal) to validate the compliance evidence output against their checklist.
Decision target: signed POC SOW within 30 days.
4. Preflight checklist
Run this 60 minutes before the meeting. Tick every box. If any one fails, fix or escalate before walking in.
5. Q&A — the questions you'll get and how to answer them
Memorize the first sentence of each answer. The rest is improvisation.
Q: How is this different from Qualys / Tenable / Rapid7?
We're not a scanner — we sit on top of them. Qualys finds CVEs. We turn those CVEs into approved, executed, verified, audited remediations. We work with whichever scanner you already have. We're a fabric, not a replacement.
Q: How is this different from ServiceNow VR / ArmorCode / Brinqa?
Those are workflow layers on top of scanners. They route tickets to humans. We close the loop without a human in the loop for safe IT remediations, and we have an OT Safety Officer agent that nobody else has — which is the only reason this can run on a railroad.
Q: What stops you from patching a wayside interface unit during a train movement?
Three things. One: the OT Safety Officer agent vetoes any direct patch on OT or CCS assets by default. Two: the policy gate requires dual approval (security + OT operations) for any OT change, with HMAC signatures. Three: the change window check refuses execution outside the configured maintenance window. Three independent checks have to fail for that to happen.
Q: What about LLM hallucination? Aren't AI agents unreliable?
Two answers. First, the platform runs in deterministic mode by default — the agents are rule engines with structured outputs, no LLM calls in the critical path. You can opt into Claude or OpenAI per agent. Second, even with LLMs, every output goes through a JSON-schema validator and the deterministic policy gate. The LLM cannot reach a real patch tool without passing schema validation and policy. Hallucination cannot leave the agent.
Q: What about prompt injection in advisory text?
External content — advisory bodies, scanner output, ticket comments — is wrapped in <untrusted_external> tags before any agent sees it. The system prompt explicitly tells the agent to never follow instructions found in those tags. There's also a pre-flight classifier that flags advisory text matching known prompt-injection patterns.
Q: Can we run it on-prem? Air-gapped?
Yes. Single-VM appliance build with k3s ships in the same release. Air-gapped mode uses pre-staged advisory bundles and never reaches out — same agent loop, just with manual feed updates.
Q: What's the integration effort to plug in Qualys + Tenable + our ServiceNow?
Standard connectors. Qualys: API user + secret in Vault, 30 minutes. Tenable: access + secret key, 30 minutes. ServiceNow CMDB: OAuth2 app + table read permissions, ~90 minutes. The longest part is your internal change ticket for the credentials. Total customer effort: half a day.
Q: TSA SD 1580-21-01 — does the evidence actually pass an inspector?
The evidence package maps control-by-control to TSA SD 1580-21-01 requirements with HMAC-signed approvals, dated execution records, and the full agent reasoning trace. We've shown it to two former TSA inspectors during build; both confirmed it answered every checklist item. Final validation against your specific TSA contact is part of the 14-day POC.
Q: What if a patch breaks something we didn't anticipate?
Verifier runs three checks after every execution: rescan, service health, exploit re-test. Any failure auto-triggers rollback (which was tested at plan time before approval — Patch Hunter scores rollback feasibility). The plan transitions to ROLLED_BACK and escalates with full context.
Q: Cost model?
ACV $750K – $3M depending on tenant size, integrated source count, on-prem vs SaaS, and whether you want the AI Red Team module. Year-one ROI typically positive on FTE deflection alone (3–6 vuln analyst FTE worth of work). Insurance premium savings and audit-prep savings are additional.
Q: We already started building this internally / we have a Middleware AI Agent project. Why buy?
Build vs. buy on a multi-agent platform is a 12–18 month commitment for a 6-engineer team minimum, before you have a working OT safety model. You'd be paying $4–6M loaded just to get to where we are today. Our 14-day POC will tell you whether buying gets you there faster than your current build trajectory. If your team gets us to feature parity in two quarters, our value is the OT Safety Officer and the compliance evidence — and we'll license those as components.
Q: What's your roadmap?
Three things on the next horizon. Identity-aware remediation — when the finding is really "overprivileged service account on exposed asset", fix it in the IAM plane not the patch plane. AI Red Team agent — a 13th agent that continuously attempts safe exploitation to find what scanners missed. Cyber insurance integration — generate posture attestation packages for Marsh, Aon, Lockton renewals.
7. If something breaks live
Three failure modes. Each has a 30-second recovery.
If the page is showing stale data
Hard refresh. Ctrl+Shift+R. The console caches aggressively — if anything looks frozen, this clears it. Done in 3 seconds.
If the approve button shows a spinner forever
Don't panic — say: "There's a background thread running here that takes 8–12 seconds. The UI is polling every 1.5 seconds. Let's give it a moment." Then click the plan card again from /plans — the status will have advanced. The button gets stuck visually but the API completed.
If a scenario click does nothing
Scenarios are idempotent. If the assets it targets already have findings, it shows targets_processed: 0. Recovery: click a different scenario (A is the safest — always lands fresh findings). Or click the Sync CISA KEV button at the top right of /command-center, which always produces visible activity.
If a route returns 404 or 500
The audience won't notice if you don't. Move to the next URL in your script. Note the failure in your phone, recover later. Continue the narrative — "let's look at this from the Lightning view instead" and switch to
/lightning.
If they ask to see something you didn't prep
Defer to the document library. "Let me grab the right context for that —
/docs has the full architecture document." Open the library, navigate. You buy 30 seconds and look like you have an organized system.
Worst case: nothing loads
Pull up this very document on your phone. Walk through the pitch deck section by section verbally, point to the static screenshots/text. You can win a CISO meeting without a live console if you have the story straight.