PROOF INTELLIGENCEFOR CRITICAL INFRASTRUCTURE
Some actions can’t be undone.
Prove them first.
Critical actions: checked against your rules, blocked if they break them.
Code migrations: what holds, what doesn’t, and the exact case that shows it, before going live.
Independent. Sealed. Verifiable without trusting us.
NEURAL · THE LANGUAGE MODEL
AI reads the code or the policy and proposes what to check. It proves nothing on its own.
SYMBOLIC · THE SOLVER
Mathematically verifies properties within defined models and assumptions.
EXAMPLE DOSSIER
PROVEN
within the stated model and assumptions
SEALED · Ed25519 + ML-DSA-65
Verify a real sealed dossier
- PROVE
- ENFORCE
- SEAL
- VERIFY
WHAT WE BELIEVE
Ironproof brings proof to the decisions that cannot be undone.
AI now writes code, moves money, and changes critical systems faster than humans can review them.
But speed without assurance is not enough.
Every critical action must be authorized before execution. Every migration must preserve the behavior that matters. Every claim must withstand independent verification.
Ironproof is building Proof Intelligence.
AI proposes, formal methods decide, and each result ships with evidence anyone can check.
Others sell trust.
We sell the right not to trust us.
WHAT IT PROVES
One product: the gate in front of critical actions.
AI agents & critical actions
For banks putting AI agents into production: keep every payment, access grant and change inside your policy, checked before it runs, and give your auditor a sealed record of each decision instead of a log.
WATCH IT DECIDEHOW IT PLUGS IN
Your system calls the gate over an HTTPS API, authenticated in both directions, before the action runs; AI agents can reach it through the MCP adapter. The credentials that execute the action live behind the gate, so skipping the check leaves nothing to act with.
Same engine, also on mandate.
The engine behind the gate also takes on three hard questions, as a mandate delivered by our team.
- 02Cloud access policiesFind the one request your access policy lets through, before an attacker does.SEE THE RUN
- 03Security fixesKnow whether a patch really closes the flaw, before you ship it.SEE THE RESULTS
- 04Translation checkSomeone else translated your code: check that it computes what the old code did, before it goes live.SEND US A TRANSLATION
WHY IRONPROOF
Faster to production. Easier to audit. Harder to break.
- 01
Improve time to market
Put AI agents, policy changes and migrations into production sooner. The checks run automatically, and the evidence is produced as you go, not assembled after the fact for the audit.
- 02
Pass the audit with evidence
Give your auditor and your regulator a sealed record of every decision that they can re-check themselves, instead of logs and screenshots.
- 03
Catch what tests miss
Get the exact input or sequence of actions that breaks your rule before production does, or the proof that none exists within the stated model.
HOW OUR PLATFORM WORKS
From your case to evidence you can re-check.
- 01
Scope
Bring one case: the policy behind an action, a security fix, a program to translate. Before anything runs, we write down exactly what will be checked.
- 02
Model
We turn it into precise logic a solver can read. Whatever we cannot model is listed as not covered, never skipped.
- 03
Prove and cross-check
A solver checks every case the model allows. A second, independent check confirms it: a second solver for policies, IBM’s compiler for COBOL, decoy fixes for C.
- 04
Seal and hand over
The proof, or the exact case that breaks the rule, sealed with a classical and a post-quantum signature. Your auditor re-runs it without us.
WATCH IT DECIDE
Four kinds of action. One gate.
Pick what the agent is trying to do. Every request is checked against the policy before it runs — and every decision, allow or block, leaves a sealed record.
INITIATOR AI agent · treasury-ops
REQUESTED ACTION
Internal transfer of $30,000, after a $35,000 wire earlier today
POLICY CHECKS
- Transfer under its own cap of $50,000
- Wires + internal transfers today under the combined cap of $60,000
- Destination account owned by the same client
Blocked. It never executes.
A rule fails. The action does not run, and the refusal is sealed with the reason.
WHAT THE SEALED RECORD CARRIES
- 01The exact action requested, and by whom
- 02The policy version in force
- 03The verdict and the rule that decided it
- 04Two signatures, one of them post-quantum, re-checkable offline
Illustrative decisions under a sample policy.
TRY TO GET PAST THE SEALED POLICY YOURSELF MORE ON THIS ACTION TYPEWHERE IT APPLIES
The rules your regulators already require.
Ironproof checks every action against them.
The actions that cannot be taken back once they run. For those, authorization stops being a setting and becomes infrastructure.
WHERE IT APPLIES — THE RULES ALREADY WRITTEN
Banking & payments
Wires, refunds and transfers · Payee changes and daily limits · Real-time payments
OSFI E-23 · OSFI B-13 · RPAA · AML programs
OTHER CRITICAL ACTIONS
Mortgage & lending
Approvals outside underwriting limits · Funding and disbursement · Changes to payout details
OSFI B-20 · PCMLTFA (mortgage sector, since Oct 2024)
Insurance
Automated claim approvals · Underwriting decisions · Payouts
OSFI E-23 · AMF guideline on AI
Online gaming
Deposit and betting limits · Payouts · Self-exclusion holds
AGCO Registrar’s Standards (Ontario)
Government
Updates to official records · Approvals and spending · Access to sensitive systems
TBS guidance on agentic AI · ITSG-33
Defence
Tasks delegated to autonomous systems · Software and configuration updates to fielded systems · Access to controlled information
DND/CAF AI Strategy · ITSG-33 · Controlled Goods Program
Healthcare
Access to patient records · Changes to orders and prescriptions · Data exports
PHIPA · Law 25 · PIPEDA
Energy & critical infrastructure
Remote setpoint and configuration changes · Emergency shutdowns and safety-instrumented logic · Water-treatment dosing and valve commands · Configuration pushed to a telecom network core
IEC 62443 · NERC CIP · IEC 61511-1 · Bill C-8, S.C. 2026, c. 9 — Critical Cyber Systems Protection Act (Royal Assent June 2026, not yet in force)
Pharmaceutical manufacturing
Electronic batch release · Changes to master batch records · Audit-trail changes
Health Canada GMP Annex 11 (GUI-0050) · FDA 21 CFR Part 11
WHY NOW
Systems are starting to act on their own.
Agents, scripts and automated pipelines now move money, change access and ship code without a person in the loop. Regulators, banks and defence departments have put dates on it. Each item links to its source.
Oct 9, 2025
The Government of Canada sets its post-quantum migration deadlines: high-priority federal systems by the end of 2031, all remaining systems by the end of 2035.
Treasury Board, SPIN on migrating to post-quantum cryptography↗Dec 3, 2025
Canada’s Cyber Centre co-signs joint guidance with CISA and allied agencies on AI in operational technology: integrating AI into “OT environments that manage essential public services” “introduces significant risks”.
Canadian Centre for Cyber Security, joint guidance↗Feb 17, 2026
Canada launches its first Defence Industrial Strategy. Its ten priority areas include digital systems, where artificial intelligence sits, and uncrewed and autonomous systems.
National Defence, Defence Industrial Strategy briefing↗Feb 2026
A paper in the proceedings of NATO’s Science and Technology Organization: agentic systems “can affect system performance or the physical environment”, and an effective agentic control regime is “essential for the trustworthiness of autonomous military and intelligence AI systems”.
NATO STO, symposium paper STO-MP-IST-210↗Mar 2, 2026
Santander and Mastercard complete Europe’s first live end-to-end payment executed by an AI agent, inside a regulated bank’s payment infrastructure.
Santander, press release↗Jun 15, 2026
Bill C-8, “An Act respecting cyber security”, receives Royal Assent. It amends the Telecommunications Act and makes consequential amendments to other Acts.
Parliament of Canada, LEGISinfo C-8↗Jul 13, 2026
OSFI publishes its bulletin on generative and agentic AI: agents that reason, plan, decide and take actions without human intervention call for least privilege and audit trails on every AI-initiated activity.
OSFI, bulletin on generative and agentic AI↗Aug 2026
FERC issues a letter order approving NERC’s proposed critical infrastructure protection standard CIP-015-2 for the bulk power system.
FERC eLibrary, docket RD26-6↗Sep 1, 2026
EY: 52% of banks have piloted agentic AI, only 16% have fully deployed use cases. “A bank cannot simply say an AI system is governed. It must be able to prove it.”
EY, How governed intelligence can scale agentic banking↗Sep 10, 2026
OSFI states that tokenized deposits are not legally distinct from traditional deposits. The existing rules apply to them.
OSFI, statement on tokenized deposits↗Sep 21, 2026
Lt.-Gen. Darcy Molstad, commander of Canadian Joint Forces Command, tells industry the military will rapidly buy more autonomous systems: 399 vendors pre-qualified for the Defence Drone Initiative, five days to answer a request for proposals.
CBC News↗Sep 22, 2026
Six Canadian banks announce they are exploring a CAD tokenized deposit solution, aiming for faster, more efficient and programmable payments.
Joint release, Newswire↗May 1, 2027
OSFI Guideline E-23 on model risk management takes effect. Its definition of a model explicitly includes AI/ML methods.
OSFI, Guideline E-23↗2030
National Defence, 2026-27 Departmental Plan: “By 2030, we aim to be a secure, agile, AI-enabled organization.”
National Defence, Departmental Plan 2026-27↗
When a system can act on its own, the decision has to be made before the action runs, not read in a log afterwards.
HOW IT WORKS
Prove. Enforce. Seal. Verify.
- 01
Prove
Before deployment, Ironproof establishes that the defined policy holds across the modeled action space.
- 02
Enforce
At runtime, every requested action is checked deterministically before execution.
- 03
Seal
Each decision is sealed — SHA3-512 digest, dual Ed25519 + ML-DSA-65 signature — binding the action, the policy digest, the acting agent and the verdict into one artifact. An allowed action is sealed before it runs; if the seal cannot be written, it does not run.
- 04
Verify
The certificate is re-checked against its sealed inputs: the same verdict must come back, or the seal is broken.
A policy proved and then broken, and the exact call site where the gate sits in front of the action, are in the technical record.
Check every action. Run what complies. Stop the rest.
Probably safe is not
provably impossible.
Proven for the actions and the policy you define, where the action can only execute through the gate. The certificate states that boundary.
START A PILOT