The US Chemical Safety Board lists completed investigations with the date each final report was released. Between 25 September 2025 and 16 September 2026 its list shows seven: a Cuisine Solutions ammonia release in Sterling, Virginia; Dow Louisiana Operations explosions in Plaquemine; a PEMEX chemical release in Deer Park, Texas; a Givaudan Sense Colour explosion in Louisville, Kentucky; the Bio-Lab fire and chemical release in Conyers, Georgia; the U.S. Steel Clairton coke oven explosion; and the Shell Polymers furnace explosion and fire in Monaca, Pennsylvania.
This post looks at two of them for what they say about automation, and about the software that increasingly sits between people and process hazards. It makes no claim that any product reduces injuries.
What the numbers say
BLS recorded 17 fatal injuries in chemical manufacturing in 2024, one of them from an explosion or fire, within 353 in manufacturing overall (Table A-1). Across all industries, BLS counted 93 fatal injuries from explosions and fires in 2024, down from 104 in 2023 (Table 2).
Two cases
Clairton, 11 August 2025. The CSB's report of 10 August 2026 says two workers died and eleven were injured. The causes included an informal practice of using pressurized water on valve seats with no written procedure, occupied buildings sited under 20 feet above coke oven gas piping and not built to withstand an explosion, and an ineffective process safety management system.
Shell Polymers, Monaca, 4 June 2025. The CSB released its final report on 16 September 2026. According to news coverage of that report, flammable cracked gas flowed backward into a furnace firebox and ignited after a process control engineer opened two isolation valves by mistake while Furnace 5 was returned to service. The coverage says Shell relied on 11 administrative procedures against backflow, a hazard it had identified as potentially fatal, although engineered controls had been offered by the furnace licensor. No injuries were reported, and Shell estimated about $95 million in property damage. Read the CSB report itself for the exact wording.
How this could have been avoided
- Where a hazard is rated potentially fatal, prefer an engineered control to another procedure. Shell's case turned on a valve sequence that people had to get right.
- Give inexperienced staff tasks with interlocks, not just instructions.
- Make the written procedure match the work. At Clairton the valve washing practice was informal.
- Site occupied buildings away from hazardous piping.
Where automation helps
Automation can keep people out of the process area, enforce valve sequences with interlocks, and raise an alarm on a leak earlier. It also moves risk into software: an interlock coded wrongly, a configuration changed without review, or a controller running tampered firmware.
What the software must get right
- Interlocks that a model cannot override. AI can advise on a startup sequence. Interlocks and safe stop sit below it.
- Human approval for bypasses and setpoint changes, with identity and reason recorded.
- Secure updates, signed and matched to an inventory, with a rollback.
- Tamper evident logs of changes, alarms and approvals, held off the controller.
- Offline operation. Plants run with the internet disconnected.
- Change control for logic. A change to safety logic goes through review like a change to a pipe.
What to do now
- List where your plant relies on an administrative control for a potentially fatal hazard and ask whether an engineered one exists.
- Obtain an SBOM for each controller, gateway and engineering workstation application.
- Compare against the CISA Known Exploited Vulnerabilities catalog.
- Keep a log of every logic change with approver.
How Safeguard fits, and where it stops
Safeguard does not run a plant or claim any safety outcome. On the supply chain side it offers SBOM generation and import, vulnerability scanning with known exploited status and reachability, firmware, container and dependency scanning, policy gates, audit logs, and on premises or air gapped deployment with offline models.
Try it yourself
Take a near miss from your own site and write the timeline: what was changed, who approved it, what logs exist. The demo below walks through an incident and postmortem.
Click through it at your own pace, or open it full screen.
Quick check
What is the difference between an administrative and an engineered control?
An administrative control relies on people following a procedure. An engineered control, such as an interlock, works by design even if a person makes a mistake.
Why should a model never override an interlock?
Interlocks are the last layer against a potentially fatal event. A suggestion engine above them can be wrong, so it must not be able to turn them off.
Sources
- CSB, completed investigations, read 18 October 2026
- CSB, final report on the Clairton Coke Works explosion, 10 August 2026
- WPXI, federal investigators release final report on Shell Polymers Monaca explosion, September 2026
- BLS, Table A-1, industry by event or exposure, 2024
- BLS, Table 2, fatal occupational injuries by event or exposure
- CISA Known Exploited Vulnerabilities catalog
