# Ot Safety First Threat Model > OT Safety-First Threat Model - Skill: `annatchijova/ot-safety-first-threat-model` (Agent Skill) - Install (CLI): `npx skillmds@latest add annatchijova/ot-safety-first-threat-model` - Raw SKILL.md: https://api.skillmd.com/api/skills/annatchijova/ot-safety-first-threat-model/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: annatchijova (https://skillmd.com/u/annatchijova) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/annatchijova/ot-safety-first-threat-model --- # OT Safety-First Threat Model Operational technology inverts the assumption every IT security instinct is built on. On a corporate network the ranking is confidentiality, integrity, availability, and the reflexes follow: patch the moment a CVE drops, scan aggressively to find exposure, force encryption and MFA everywhere, reboot to apply. In OT — the PLCs, RTUs, HMIs, and safety systems that run a turbine, a pipeline, a water plant, a factory line — the ranking is **safety and availability first**, and those same reflexes become the threat. A vuln scan can crash a fragile controller that has an uptime measured in years. A forced patch can require a process shutdown that costs a week and a restart that is itself dangerous. An "encrypt everything" mandate can add latency that violates a real-time control loop. The wrong security action does not leak data; it stops a physical process or defeats a safety function, and the consequence is measured in downtime, environmental release, or lives. Two symmetric failure modes: - **IT reflexes imported blind.** Treating the plant floor like an enterprise LAN — scanning it, patching on the IT cadence, deploying agents onto devices that were never built to run them — and inducing the outage you were trying to prevent. The security team becomes the incident. - **"It's air-gapped, it's fine."** The mirror error: assuming physical isolation that no longer exists (a historian bridged to the corporate network, a vendor's remote- maintenance modem, a USB workflow, IT/OT convergence) and therefore doing nothing, leaving a flat, unauthenticated, unpatchable network wide open once the gap is crossed. The through-line: **model consequence in the physical world, and let availability and safety, not confidentiality, set the priority.** Every control is judged first by what it does to the process. Composes with the library: - **assume-breach-modeling** — the blast radius of an OT compromise is kinetic; "what does this foothold reach" ends at a physical actuator, not a database - **irreversible-action-gate** — a control command (open a valve, change a setpoint, stop a safety system) is a one-way, physical action; the gate is not a UX nicety here, it is safety engineering - **honest-degradation** — an unpatchable, end-of-life controller that cannot be fixed is a documented WARN with compensating controls, not a pretended PASS - **dual-use-behavior-adjudication** — on an OT network, an engineering workstation issuing control commands is normal; malice is context (unexpected setpoint, off-hours, from the wrong host), not the command itself - **detection-engineering** — OT detection is passive and protocol-aware (span ports, not agents; Modbus/DNP3/S7 semantics), because active probing is itself a hazard --- ## Step 1 — Rank the impact in the physical world, safety at the top Before any control decision, state the consequence hierarchy for *this* system, because it determines what "secure" even means: - **Safety** — can a compromise or a security action cause physical harm (release, explosion, equipment destruction, injury)? A safety-instrumented system (SIS) that a control defeats or delays is the top of the risk stack, above any data concern. - **Availability** — can it stop the process? Downtime in continuous processes is not an inconvenience; a restart may be slow, costly, or hazardous in itself. - **Integrity** — a manipulated setpoint or spoofed sensor reading that drives the process to an unsafe state (the Stuxnet / Triton class), where the operator sees "normal" while reality diverges. - **Confidentiality** — last, not first. Process data matters, but not at the cost of the three above. An assessment that opens with "we should encrypt the historian" before it has characterized the safety and availability impact has imported the wrong ranking. --- ## Step 2 — Test every proposed control against availability and safety first For each IT-style control on the table, ask what it does to the process *before* asking what it protects: - **Patching** — does applying it require a shutdown? Is the vendor's patch validated for this device and firmware? Is the controller so fragile or end-of-life that the patch risk exceeds the vuln risk? Patch cadence in OT is driven by maintenance windows and vendor validation, not by CVE date. - **Scanning / active discovery** — can this device tolerate a scan, or will an unexpected packet crash a decades-old TCP stack? Prefer passive discovery; active probing is a hazard until proven safe on that exact model. - **Agents / encryption / MFA** — will the endpoint even run an agent? Does encryption break a real-time loop's timing budget? Does MFA on an HMI block an operator during an emergency when seconds matter? A control that impedes emergency operation is a safety regression. A control that reduces a confidentiality risk while raising a safety or availability risk is usually the wrong trade in OT. Name that trade explicitly. --- ## Step 3 — Verify the isolation you are assuming actually holds The "air gap" is the most over-claimed control in OT. Do not assume it; map it: - **Enumerate every crossing** — IT/OT firewalls and their real rules, historians and data diodes, jump hosts, vendor remote-access (cellular modems, TeamViewer, dial-up still exists), USB and laptop workflows, the engineering workstation that touches both worlds. IT/OT convergence has quietly bridged most "gaps." - **Treat the crossing as the trust boundary** — the corporate network is untrusted from the OT side; the engineering workstation and the historian are the classic pivot points (assume-breach-modeling). Once inside, OT networks are typically flat, unauthenticated, and unencrypted by design — so a single crossing is often domain-wide on the process network. - **If the gap is real, defend the few crossings hard rather than the many endpoints you cannot patch.** If it is not real, say so — a believed-but-false air gap is the worst of both worlds. --- ## Step 4 — Build compensating controls around what you cannot fix Much of OT cannot be patched, upgraded, or hardened without unacceptable process risk. Honest degradation is the posture, not a failure: - **Segment and monitor instead of patch** — network zones/conduits (the Purdue model / IEC 62443 idea), tight allow-lists at the boundary, and *passive* protocol-aware monitoring that raises no traffic on the wire. - **Protect the safety layer independently** — the SIS should be separated from the basic process control so that a compromise of the controllable layer cannot also defeat the system that would stop a runaway (the Triton lesson). - **Document the residual risk as a WARN with a reason** — "controller X is end-of-life, unpatchable; compensating controls are segmentation + passive monitoring; residual risk accepted by process owner" is honest and auditable. A pretended-clean status on an unpatchable device is the dangerous lie. --- ## Step 5 — Detect and respond without becoming the incident Detection and IR in OT carry the same inversion: - **Passive, protocol-aware detection** — baseline the process traffic (Modbus, DNP3, S7, EtherNet/IP) and alert on the semantics: an unexpected setpoint change, a firmware download to a PLC, an engineering command from a host that never issues one (dual-use-behavior-adjudication, physical edition). No active agents on the process devices. - **Containment cannot default to "isolate and reboot"** — pulling a controller offline or cutting the HMI may be more dangerous than the intrusion. Containment decisions run through the process/safety engineers, weighing the physical consequence (containment-under-uncertainty with a kinetic blast radius). The plant may need to run degraded, not stop. --- ## The one-line test If your OT recommendation would be uncritically correct on a corporate laptop — patch it now, scan the subnet, force encryption, isolate and reboot — you have not threat-modeled OT; you have pointed IT reflexes at a physical process. Re-rank with safety and availability on top, and test each control by what it does to the process before what it protects.