Industries / Industrial equipment and services / IoT and industrial automation / Industrial controls and PLCs
Safety controllers
Safety controllers are the separate devices and programs that stop a machine or bring a plant to a safe state when something goes wrong, built, certified and programmed to their own standards.
7
Milestones achieved, from 1986 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1986
Three processors vote
Triconex shipped its first Tricon in 1986, a safety controller that runs the same logic on three independent channels and acts on two-out-of-three votes, so that one failed part neither disables the safety function nor trips the plant by itself. Before it, one faulty relay or controller could defeat a whole shutdown circuit, a 2025 Schneider Electric account says. In January 2002 the US Nuclear Regulatory Commission approved Tricon Version 9 for safety systems in nuclear power plants, in Triconex's words the first triple modular redundant (TMR) system the commission had qualified.
Credit
Triconex, founded in September 1983 by Jon Wimer in Santa Ana, California
1987
A relay for stop buttons
Pilz launched the PNOZ in 1987, a compact emergency stop relay that makes a machine stop reliably when someone presses the stop button, and calls it the first safety relay. It was smaller and easier to use than the conventional relay circuit it replaced, and it had passed a certified type examination. Pilz later sold the PSS 3000, which it calls the world's first safe programmable controller.
Credit
Pilz, then run by Renate Pilz
1995
Why control systems fail
The UK Health and Safety Executive (HSE) analysed 34 incidents in which control systems failed and published the results in "Out of control" in 1995. It traced 44% of the primary causes to the specification, 20% to changes made after the system was commissioned, 15% each to design and to operation and maintenance, and 6% to installation. The 2003 second edition kept the 1995 analysis unchanged.
LessonMost failures began before the system first ran; specification, design and installation together account for 65% of primary causes in HSE's table.
Credit
HSE
1998
One standard for safety electronics
The International Electrotechnical Commission (IEC) published the first edition of IEC 61508, "Functional safety of electrical/electronic/programmable electronic safety-related systems," in 1998. Functional safety is safety that depends on a system working correctly. The standard defines a safety lifecycle, the steps from specifying a safety system to retiring it, and was written to be the base for standards in particular sectors. Its second edition followed on 30 April 2010.
Credit
IEC; HSE's Electrical and Control Systems Unit "strongly contributed," by HSE's account
1998
Safety signals share the cable
PROFIBUS & PROFINET International (PI) began PROFIsafe in autumn 1998, a way to send safety signals over the same PROFIBUS and PROFINET networks as ordinary data, so that a second, hard-wired safety circuit was no longer needed. It introduced the "black channel" principle: the safety layer trusts nothing about the network and catches lost, repeated, out-of-order, corrupted or late messages at each end with sequence numbers, timeouts and a 24-bit checksum. PI counted 410,000 PROFIsafe nodes by the end of 2007. In August 2025 Tobey Strauch, writing in Control Design, put PROFIsafe's share of safety networks at 80 to 90%, without naming a source, and said CIP Safety and Safety over EtherCAT followed the same principle.
Credit
PI
2011
Machine categories retired
The European standard EN 954-1 graded the safety circuits of machines in categories set by how they were built, such as whether they had redundant parts. EN ISO 13849-1, accepted as a European standard in 2006, replaced the categories with performance levels that also count the probability of a dangerous failure, diagnostics and common-cause failures, and that cover programmable safety devices and their software. EN 954-1 was due to lapse on 30 November 2009; the date was moved to 31 December 2011, after which it no longer showed that a machine met the EU Machinery Directive.
LessonA circuit's structure alone did not say how often it would fail or how its software was checked, and moving the trade took a two-year extension.
Credit
ISO, with EN ISO 13849-1 adopted in Europe; Thomas Maier of UL (Underwriters Laboratories) explained the change
2025
A safety program proved
GSI, Germany's heavy-ion research centre in Darmstadt, had the PLC program of the Personnel Access System for its FAIR accelerator, which keeps people out of areas exposed to particle beams, formally verified. CERN's tool PLCverif checked the program against written requirements with model checkers, programs that explore every state a design can reach. GSI wrote the requirements, one CERN team the program and a separate CERN team the proof; the discrepancies found were fixed, and PLCverif gained models of Siemens functions whose code is hidden, such as the ESTOP1 emergency stop block. The authors call it "the first time an organization trusted another to formally verify a complete, real-world, safety-critical PLC project."
LessonFunctions whose code the maker hides had to be modelled from the maker's documentation and from simulation before the whole program could be proved.
Credit
Ignacio D. Lopez-Miguel (TU Wien), Borja Fernández Adiego (CERN), Matias Salinas and Christine Betz (GSI)
Watch for
AI drafts safety code
Neuron Automation says the Transforma extension of its Neuron Smart Engineer programming environment generates Structured Text code from structured safety requirements, converts it to function block diagrams, writes test cases and helps find errors. The company says the environment meets IEC 61508 requirements up to SIL 3 (safety integrity level 3) and can run with language models offline.
Achieved whenan outside party, such as a certifier or a customer, publishes how often the generated code and tests were correct on a set of real safety functions.
Claimed by
Neuron Automation, July 2026
IEC 61508's third edition
IEC's catalogue lists edition 3.0 of IEC 61508-1 as under development, at the stage where the result of the vote on its draft is being prepared, with a forecast publication date of 22 October 2026, which has passed.
Achieved whenIEC publishes the third edition.
Claimed by
IEC, catalogue entry read in November 2026
Rules for AI in safety
ISO and IEC's joint AI committee, ISO/IEC JTC 1/SC 42, is drafting ISO/IEC TS 22440, requirements for AI inside safety-related control systems, AI that supports them, and AI tools used to develop them. It follows ISO/IEC TR 5469, a 73-page report from January 2024 that only describes risk factors and methods. Part 1 is a committee draft on ISO's site, and Jody Nelson of the consultancy SRES wrote that it is not expected before early to mid-2027.
Achieved whenISO and IEC publish Part 1.
Claimed by
ISO/IEC JTC 1/SC 42; Jody Nelson (SRES), October 2026
Open
01
Prove AI-written safety logic
A public benchmark of machine safety functions, such as an emergency stop, a guard-door interlock, two-hand control and light-curtain muting, each written as a plain specification. AI models write each program in Structured Text with standard safety function blocks, and every program is checked two ways: a model checker proves it against the specification, and a simulation injects faults such as a stuck contact, a broken wire or two channels that disagree.
The briefCredit · unclaimed02
Recount why control systems fail
A public, coded set of incidents since 1995 in which a control system or safety controller contributed, taken from published investigation reports, with each incident assigned to HSE's lifecycle phases by two people, to test whether the specification still causes the largest share of failures.
The briefCredit · unclaimed
Open 01 · the brief
Prove AI-written safety logic
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark of machine safety functions, such as an emergency stop, a guard-door interlock, two-hand control and light-curtain muting, each written as a plain specification. AI models write each program in Structured Text with standard safety function blocks, and every program is checked two ways: a model checker proves it against the specification, and a simulation injects faults such as a stuck contact, a broken wire or two channels that disagree.
Why it is still open
PLCWorld, an October 2026 benchmark by Yunji Kim and colleagues, counts safety violations by AI-written programs, such as moving without an interlock or colliding, in 100 simulated motion and material-handling tasks, but not how safety functions behave when a part fails. A team including Andreas Schwung at Fachhochschule Südwestfalen scored the "safety compliance" of AI-written code on 25 use cases with an AI judge and human review, not with proofs, and Neuron's generator above has no published test.
Built with what exists
PLCverif, CERN's verifier for PLC programs (open-source version on GitLab as plcverif-oss), which since the GSI project models Siemens safety blocks such as ESTOP1; PLCopen's safety function block specification, which the German institute BGIA reviewed against IEC 61508; the OpenPLC runtime or CODESYS to run programs against simulated inputs; AI coding agents to write programs and the properties to prove (see the Note "Proven in four weeks").
First step
Write the emergency-stop function as five properties, hand-write a reference program, get PLCverif to prove it, then score one model's version against the same properties.
Done looks like
30 safety functions, programs from at least three AI models, and for each program a published proof result or counterexample plus the outcome of at least ten injected faults, with the specifications, programs and scripts released.
Open 02 · the brief
Recount why control systems fail
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, coded set of incidents since 1995 in which a control system or safety controller contributed, taken from published investigation reports, with each incident assigned to HSE's lifecycle phases by two people, to test whether the specification still causes the largest share of failures.
Why it is still open
In 2002 HSE planned a successor to "Out of control" that would "contain results for a greater number of incidents," but the 2003 second edition says its analysis "remains unchanged from the first (1995) edition." HSE also notes that its classification of each incident is a judgement on which "others may have differing views." Searches found no published recount.
Built with what exists
published investigation reports from bodies such as the US Chemical Safety Board and HSE; the eight cause categories in "Out of control"; AI models that read long reports, such as Claude or Gemini, for a first coding pass (see the Note "The jobs agents repeat"); two human coders, with agreement measured by Cohen's kappa, a standard score for how often two raters agree beyond chance.
First step
Give an AI model the 34 incident summaries in Appendix one of "Out of control," have it assign each a primary cause, and compare its calls with HSE's.
Done looks like
At least 150 incidents coded by two people with a kappa of at least 0.7, the share of primary causes per phase published beside HSE's 1995 shares, and the data released under an open licence.