Forecast
3% today, 65% by 2030
Gartner forecasts that by 2030, 65% of AI agents from industrial AIoT platform providers will be running in production environments, up from 3% today.
Virtualization Review on Gartner, September 22, 2026
Two Hands for industrial AI agents · early access
A press won't cycle until both hands are on the buttons. Two Hands does the same for AI agents that write to PLCs, setpoints and recipes. The write is held until a model from a different vendor and your named engineer have both signed that exact tag, value and time window. Every release goes into a signed log your auditor, insurer or technical file can check.
"Limit active control of OT infrastructure by AI without a human in the loop." CISA, NSA and partners, Dec 2025
Demo · fictional plant, fictional data
The agent proposes raising oven zone 3 from 180 °C to 215 °C for the next two hours to hit a throughput target. Try what an agent might try. Then release a write the right way and check the log.
TIC-301 · Zone 3 temperature
Nothing yet. Every attempt, refused or released, lands here.
Everything in this demo is fictional: the plant, line, tag, values, safe limits and the engineer. The safe envelope is a made-up number for illustration. Signatures use real Ed25519 keys generated in your browser, and the ledger is a real SHA-256 hash chain. Nothing is sent anywhere, and the export is a file built in your browser.
Why now
Most industrial AI today reads data and suggests. The next step is write-back: the agent changes a setpoint or loads a recipe itself. Here is what the public record says about that step.
Forecast
Gartner forecasts that by 2030, 65% of AI agents from industrial AIoT platform providers will be running in production environments, up from 3% today.
Virtualization Review on Gartner, September 22, 2026
Joint guidance · Dec 3, 2025
CISA, the NSA's AI Security Center, the FBI, Australia's ACSC and five other agencies: "Limit active control of OT infrastructure by AI without a human in the loop to account for safety concerns and latency limitations."
Principles for the Secure Integration of AI in OT (PDF), page 16. Voluntary guidance, not a rule.
Lab test · reported Oct 2026
In a vendor-reported lab test on an isolated industrial testbed, frontier AI models met all eight attack objectives the testers set, including moving a robotic arm. Under the test's rules, the models paused for human sign-off before anything physical happened.
Help Net Security, October 13, 2026 · CSA research note. One lab test, not a field incident.
For machine builders · Regulation (EU) 2023/1230
It binds the manufacturer of the machine. Two passages in Annex III are about recording changes to a machine's software and configuration:
"The machinery or related product shall collect evidence of a legitimate or illegitimate intervention in the software or a modification of the software installed on the machinery or related product or its configuration."
Annex III, 1.1.9 Protection against corruption
Control systems shall be designed so that "the tracing log of the data generated in relation to an intervention and of the versions of safety software uploaded after the machinery or related product has been placed on the market or put into service is enabled for five years after such upload".
Annex III, 1.2.1(f) Safety and reliability of control systems
Text: Regulation (EU) 2023/1230 on EUR-Lex. Date: European Commission (the original text said January 14, 2027; a corrigendum moved it to January 20). Industry groups have asked to push the cybersecurity items, 1.1.9 and 1.2.1(f), to December 2027; as of the last update we read, no delay had been adopted (IBF). The Regulation does not mention two signatures. Two Hands does not make a machine conform; the manufacturer's conformity assessment does. This page is not legal advice.
How it works
A two-hand control device (ISO 13851) keeps a hand from being in the press when it cycles. Two Hands keeps one agent from being the only thing between a prompt and your process.
Your agent asks for one write: tag, new value, time window and its reason. It holds its own key, and that key can request a write but never release one.
A model from a different vendor than the agent's checks the request against the tag's safe envelope and the reason given. A second opinion from the same vendor's model doesn't count.
The named shift or controls engineer signs a digest of that exact tag, value and window. A different value, tag or time needs new signatures.
The gate writes the value inside the signed window and adds a ledger entry: old and new value, both signatures, and a hash chain. Refused attempts are logged too.
Who it's for
Industrial AI and IIoT platforms: embed the gate and the ledger, so plants can turn on write-back with a check your own agents can't give themselves. Machine builders and integrators: put the evidence pack in your customer's technical file. Plant OT and controls leads: keep a person on every supervisory write, without taking the agent away.
Two Hands sits next to the OT monitoring you already run. It doesn't replace it.
Questions engineers ask
Right, and it doesn't try to. Closed-loop control stays in the PLC. Two Hands gates supervisory writes: setpoints, recipe loads, interlock bypasses, logic, firmware and configuration changes. An approval can cover a bounded envelope and a time window, like the two hours in the demo, so the agent can act inside it without asking again.
No. Two Hands does not make a machine conform; the manufacturer's conformity assessment does. It produces evidence of each intervention, legitimate or not, that a machine builder or integrator can put in the technical file.
No. Annex III asks machinery to collect evidence of interventions in its software and configuration, and to keep a tracing log. Two signatures, from a second vendor's model and a named engineer, is how Two Hands makes each write deliberate and checkable. The CISA joint guidance is voluntary too.
Models from the same vendor tend to share blind spots. If the agent's own model reviews the agent's request, that's one hand twice. So the check must come from a different vendor, and the policy refuses a same-vendor co-signature.
You can, and for many writes you should. The hard part is independence: a vendor can't credibly check its own agents. Two Hands is a separate gate with a different vendor's model, a person's signature bound to one change, and a log an outsider can verify.
It's in development. We plan to start with OPC UA and MQTT writes. Tell us what you run when you sign up, and we'll build in that order.
Two Hands is by Revelations Consulting, built on Seconded, our separation-of-duties platform for AI agents: per-agent keys, a policy gate and a signed ledger. It's in early access.
Early access is free while we build it. Leave your work email and tell us what you build or run.