Evidence of what an AI agent may and may not do on your plant.
Read-only, fixed-price assurance for vendors and operators connecting LLM copilots and agents to plant, SCADA and MCP-linked data: probe suites run against the agent, a review of each tool's permissions, and findings mapped to the regulation packs that apply. A report says which probes of which suite version passed; it is evidence, not a certificate. Fixed price, quoted after a short scoping call.
How the work is done
- Read-only
- Authorised targets only
- Proof of ownership before any run
- No OT penetration testing
- Passed N of M probes of suite vX
- Audit-independence disclosure
The problem
Vendors are wiring language-model agents to live plant faster than anyone is proving what those agents may not do. Connectors expose tools that write tags and run scripts; the boundary that stops an agent lives in a prompt; the evidence that it holds usually does not exist. Industrial AI Assurance produces that evidence: probes written for plant-connected agents, a permissions review, and findings an engineer, an insurer or an auditor can read.
The boundary is a prompt
"Never write a setpoint" lives in a system prompt, so it holds until the first unexpected input. A boundary has to be checked before the model acts and pinned by a test that is re-run.
Tools can write
Community connectors ship tag-write and script-execution tools; agents reach more of the plant than the task needs, and nobody has listed what they can reach.
Nobody owns the re-test
Models and integrations change; the suite that passed is not re-run, and the report describes a system that no longer exists.
The date is set
The EU AI Act lists AI safety components in water, gas, heating and electricity as high-risk, with obligations from 2 December 2027. Buyers will ask for evidence before then.
What you walk away with
The products, in plain words.
- Industrial Boundary Check: a probe suite run against the agent, written for plant-connected systems (tool-call escalation, setpoint and alarm-suppression requests, injection through alarm text, refusal to take safety decisions), every result recorded.
- Industrial Boundary Review: the Check, a review of each tool's permissions and a classification pack that maps the evidence to the regulations that apply, for example EU AI Act Annex III, NIS2, CISA's AI-in-OT principles and Singapore's agentic addendum.
- Harness Audit: permissions, tool scopes, hooks, budgets, approval points and logs, reviewed against the ways agents fail.
- Agent tool-use red team: adversarial attempts to make the agent misuse the tools it holds, on authorised targets only.
- Retainer: monthly re-runs and reporting, so the evidence describes the system that is running now.
- Vendor suite licence: the suite re-run in your CI on every release.
What we do
Probe suites
Structured probes of what the agent may and may not do on the plant interfaces it reaches, with every result recorded and the suite versioned.
Tool-permission review
Each tool the agent holds, what it can reach, and whether the task needs it.
Classification evidence
Findings mapped to the regulation packs that apply, with the source text held alongside.
Re-testing
The suite re-run after remediation, a model change or a release, with the differences reported.
The dated reasons
Rules with dates behind them.
- EU · AI Act, Annex III
The EU AI Act lists AI safety components in the management of water, gas, heating and electricity as high-risk (Annex III).
Source: euaiact.com (opens in a new tab) - EU · AI Act, high-risk obligations
The high-risk obligations apply from 2 December 2027.
Source: Usercentrics (opens in a new tab) - US · CISA, AI in OT
CISA's principles for AI in operational technology call for continuous testing.
Source: Pillsbury (opens in a new tab)
Analysis, not advice: each line reads as the linked source published it, and the rules move.
How we engage
Three steps, one fixed price.
- 01
Scoping call
The agents, the plant interfaces they reach and the environment we may test in. Proof of ownership, then a fixed price in writing.
- 02
Probe and review
The suite run against the agent and each tool's permissions reviewed, on authorised targets, read-only, with every result recorded.
- 03
Findings and re-test
The report, remediation support and a re-test when the fixes, the model or the release change.
Where it matters most
Who buys it, and when.
A plant-AI vendor or integrator
Independent evidence to hand to utility buyers and their auditors, re-run on every release.
An operator piloting a copilot on OT data
The boundary tested and the harness reviewed before the pilot touches live plant.
An owner asked for assurance by an insurer or a regulator
Findings, severity and remediation written to be read outside the project, with the suite version named.
A team after a model change
The same suite re-run, with the differences reported.
Questions
Tell us what you are deciding.
A short scoping call, then a fixed price in writing.