← Back to writing
Critical InfrastructureSupply ChainAINational SecurityArchitecture

The Undeclared Dependency: Every Technical Supply Chain Has a Strait of Hormuz

15 min read
The Undeclared Dependency: Every Technical Supply Chain Has a Strait of Hormuz

The TL;DR is that AI cannot secure a supply chain you have not described. Australia's fuel stocks, your cloud builds, and the defence industrial base are all exposed through the same weakness, dependencies nobody wrote down, and the unglamorous act of declaring the graph and then verifying it continuously is the actual security control. The clever model you point at the problem comes a distant second.

This article argues that the security move everyone skips is the boring one. Declare every edge in the graph. Verify the provenance of each edge. Treat anything undeclared as attack surface. Fuel already does this: days of cover for diesel is a query against a graph someone declared decades ago. Software can do it too, with a bill of materials and build provenance, if we stop treating the paperwork as the point.

When the strait closed in March, which of your systems could tell you, within a day, everything downstream that would break? And if the answer is none, what exactly is your AI meant to analyse?

Introduction

It is early March 2026, and you own supply-chain risk at an Australian critical infrastructure operator. The Strait of Hormuz has just gone dark. War broke out around 28 February, and within days, by about 4 March, the world's most important oil chokepoint was declared shut for the month ahead (International Energy Agency, 2026). Global oil supply falls by 10.1 million barrels per day to 97 million barrels per day, which the International Energy Agency calls the largest disruption in history (International Energy Agency, 2026, p. 1). Physical crude touches a record near $150 USD a barrel. North Sea Dated trades around $130, roughly $60 above where it sat before the war started (International Energy Agency, 2026, p. 1).

Within a fortnight, your board asks you two questions that sound unrelated. On Monday, they want to know how exposed the organisation is to the fuel contracts now caught behind the Strait. On Thursday, after a vendor advisory lands, they want to know whether a compromised build dependency is sitting in your software estate. You go home on Thursday night and realise the uncomfortable truth. Those were not two questions. They were the same question, asked of two different graphs, and you can answer one of them.

Twenty-Nine Days of Diesel

Start with fuel, because fuel is the one place Australia can actually answer the question.

When the Strait shut, the responsible minister could stand up and give a number. Australia held roughly 34 days of diesel cover, 36 days of petrol, and 32 days of jet fuel, with the two remaining refineries, Lytton and Geelong, supplied at least until May (Argus Media, 2026). Cover sat in the low-to-mid thirties and fell as the crisis wore on, with emergency projections dropping toward the mid-twenties for diesel and lower still for jet (Hepburn, 2026). Earlier commentary had put the working range somewhere around 29 to 36 days depending on the fuel and the assumptions. The point is not the precise figure. The point is that a precise figure exists at all.

Sit with how exposed that supply chain is. Australia imports roughly 90 per cent of the refined fuel it burns (Hepburn, 2026). Around 83 per cent of those maritime imports arrive through the Indonesian straits, Malacca, Lombok and Sunda, a second set of chokepoints stacked behind the first (Lockhorst, 2026). Darwin sits less than four days' sailing from Singapore (Lockhorst, 2026). In International Energy Agency terms, Australia held about 50 days of stock against a 90-day obligation at the end of 2025, the only member of the agency out of compliance. The minimum stockholding obligation for diesel importers had already been lifted from 20 to 32 days of cover from July 2024 (Norton Rose Fulbright, 2024). When the crisis hit, the government stood up a reactive Fuel Supply Taskforce and reportedly relaxed the minimum obligation by 20 per cent to keep product moving. The 2026-27 Budget then put $14.8 billion behind a fuel resilience package that lifts diesel and jet reserves toward 50 days of cover (Commonwealth of Australia, 2026).

Here is the contrarian bit. None of that is a failure story. It is a legibility story. Fuel is the one supply chain where Australia can have a precise national argument about resilience, because someone declared the graph decades ago. Days of cover is a query. Obligations are set in days. The statistics are published. When the worst happened, the country could reach for a number, argue about it, and act on it inside a day. Nobody in this country can have that argument about their software estate, because nobody declared that graph. That is the whole essay in one contrast.

Slide the days forward. The obligation line never moves; the tanks do. The gap between them is the vulnerability.

The Same Graph in Different Clothes

Now trace the fuel chain end to end and look at the shape underneath. Crude field, to tanker, to refinery, to import terminal, to depot. It is a directed graph. Each node depends on the one upstream, and the whole thing only holds if every edge is real and every edge is known.

Point that same shape at software and it is identical. Upstream repository, to package registry, to build, to artefact store, to deployment. Point it at hardware, and it is the same again. Fabrication plant, to component, to board, to platform. Three industries, one picture.

The concept that ties them together, named once, plainly, is this: verified provenance over an explicit dependency graph. Declare every edge. Verify where each edge came from. Treat anything undeclared as attack surface. That is the control. Everything else is decoration.

Software already has the vocabulary for it; we just do not use it seriously. A software bill of materials lists your components, and its Dependency Relationship field, which characterises how an upstream component is included in the software you ship, is literally the graph edge written down (National Telecommunications and Information Administration, 2021). Supply-chain Levels for Software Artifacts (SLSA) adds provenance, a signed attestation of who built an artefact, from what source, producing what digest, graded from build level 0 to build level 3 (Open Source Security Foundation, 2023). And hermetic builds are the enforcement move. A hermetic build declares every transitive dependency up front with immutable references and runs with no network access, which means the declared graph becomes the only graph the build can physically reach. In the specification's own words, hermeticity guarantees that the dependencies are complete (Open Source Security Foundation, 2021). That last line is the entire argument. Completeness is not a nice-to-have. It is the difference between a graph you can secure and a graph that lies to you.

Two supply chains, one graph. The edge you did not declare is the one that fails you.

Shelfware With a Digital Signature

So if the bill of materials is the graph, why has it not saved anyone?

Because most bills of materials are generated to satisfy a contract clause, attached to the deliverable, and never queried again. They are a compliance artefact cosplaying as a security control. The document was never the point. The graph was the point, and the graph only earns its keep if you can query it the moment a vulnerability lands.

Log4Shell is the canonical test, and Australia was in the room for it. Just before Christmas 2021, a joint advisory co-signed by the Australian Cyber Security Centre alongside its partners went out on the Log4j vulnerabilities (Cybersecurity and Infrastructure Security Agency et al., 2021). The United States Cyber Safety Review Board then studied the event and found Log4j to be an endemic vulnerability, with vulnerable instances expected to persist for a decade or longer (Cyber Safety Review Board, 2022). Read the board's report closely, and the bottleneck was not the fix. It was finding the instances. Organisations could not answer a simple question: where is this thing in my estate, because they had never declared the graph. If every one of them had held a live, queryable bill of materials, Log4Shell would have been a bad week, not a decade-long tail.

The pattern only got sharper from there. SolarWinds put a nation-state implant into a trusted update (Cybersecurity and Infrastructure Security Agency, 2020). 3CX became, in the investigators' words, the first time a software supply-chain attack led to another software supply-chain attack, one trusted vendor compromised through another (Mandiant, 2023). Then XZ Utils, where a patient maintainer spent years earning trust before slipping a backdoor into a core Linux compression library aimed at the SSH daemon (National Vulnerability Database, 2024). Every one of those attacks lived on an edge nobody was watching, because nobody had declared it worth watching.

The fix that turns declaration from paperwork into a control is the SLSA provenance and hermetic-build combination above. Sign what you build, attest where it came from, and make the build unable to reach anything you did not declare. And yes, before my architecture friends email me, SPDX versus CycloneDX is a religious war, and for now, any is better than none.

Lessons From a War of Seams

If you want to see undeclared dependencies weaponised at scale, look at Ukraine.

On 13 February 2026, Google's Threat Intelligence Group published its assessment of threats to the defence industrial base (Google Threat Intelligence Group, 2026). Two clusters in that report are worth keeping firmly apart. One group, tracked as UNC5976, ran malicious remote-desktop connection files and stood up hundreds of domains spoofing defence contractors across nine countries. A separate group, UNC4221, ran a ClickFix social-engineering chain, tricking a user into pasting a malicious command, against a likely Ukrainian military entity (Google Threat Intelligence Group, 2026). Different tradecraft, different actors, same target set: the suppliers, not the front line. The report also notes that Russia-nexus activity concentrates on drone and unmanned-systems firms, while China-nexus activity is the highest by volume (Google Threat Intelligence Group, 2026).

Notice where all of it lands. Not on the hardened node, the frontal assault on the well-defended target. On the seams, the trust between suppliers. Sandworm sat inside Kyivstar, Ukraine's largest mobile operator, for around six months before wiping it (Claburn, 2024). On the morning of the 2022 invasion, an attack on the Viasat KA-SAT network bricked modems across Europe, and the spillover knocked out remote monitoring of roughly 5,800 wind turbines in Germany (Guerrero-Saade & Mendoza, 2022). That edge sat in nobody's risk register. A wind fleet depended on a satellite network in a war on another continent, and its operators found out on the day. NotPetya is the pre-war canon of the same lesson, a 'trojanised' update to Ukrainian tax software that escaped its blast radius and did billions in damage worldwide (Greenberg, 2018).

There is a tempo lesson stacked on top. Ukraine iterates drone designs in weeks, feeding what happens on the battlefield straight back to manufacturers, against traditional procurement cycles measured in years (Bondar, 2025). Russia, for its part, is industrialising drone autonomy toward a six-figure annual output by the end of the decade (Bondar, 2026). The side that can redraw its dependency graph fastest wins. A graph you cannot redraw quickly is a graph you have already lost.

And here is the contrarian turn most threat models miss. The European Union's cybersecurity agency found that the regional threat picture is dominated by opportunistic criminals whose targeting is often triggered by geopolitical events (European Union Agency for Cybersecurity, 2025). In a crisis, your most likely exploiter is not the patient nation-state you built your threat model around. It is the opportunist monetising the chaos at your weakest undeclared seam. So threat-model the graph's weakest edge, not the scariest actor. The grub who finds the open door beats the genius who was going to pick the lock.

Where AI Actually Earns Its Keep

I promised this would be helpful, not just contrarian, so here is where AI genuinely pulls its weight. Every one of these sits after declaration, not instead of it.

  • Entity resolution across supplier names. The same company shows up three ways across your contracts, your manifests, and the corporate registry. Resolving those into one entity is exactly the discipline I wrote about for financial-crime monitoring in The 80,000-Entity Wave, and it is where AI shines.

  • Extracting edges from unstructured documents. Procurement contracts, statements of work, and vendor questionnaires are full of dependency edges written in prose. Pulling them into the graph is a language task, and a good one for a model.

  • Anomaly detection over declared provenance. Once you have signed attestations flowing, a model can flag the build that came from the wrong source or the artefact whose digest does not match its history.

  • Simulation and wargaming. Drawdown models, days-of-cover projections, cascade analysis. Point a model at a declared graph, and it will help you rehearse the failure before it happens.

  • Prioritising what to map next. No one maps the whole estate at once. A model can tell you which undeclared corner is the riskiest place to start.

Now the part the vendors will not like. A whole product category sells AI-powered supply-chain-risk detection over an estate nobody has described. That is selling radar for airspace you never surveyed. Detection coverage is a function of declared inventory. It has a ceiling, and the ceiling is set by how much of your graph you have actually declared. That is not a statistic I am quoting from anyone's brochure. It is the argument of this article, and it rests on two hard facts: SLSA's own line that hermeticity is what guarantees the dependencies are complete (Open Source Security Foundation, 2021), and the Cyber Safety Review Board's finding that the thing which made Log4Shell endemic was the inability to find the instances (Cyber Safety Review Board, 2022). AI over an undeclared graph inherits that same blindness. It cannot flag an edge it was never shown.

There is a governance thread here too. In February 2022, the United States Department of Defense issued its continuous authorisation memo, making ongoing monitoring, active cyber defence, and a secure software supply chain the standing price of staying authorised, backed by a live dashboard rather than an annual sign-off (McKeown, 2022). Australia has no codified equivalent. The Information Security Manual model is point-in-time authorisation with ongoing monitoring and periodic reassessment (Australian Signals Directorate, 2026). A once-a-year attestation over a living graph is a static authority to operate wearing a hi-vis vest. Days of cover works precisely because it is a continuous query, not an annual audit. Your software graph deserves the same treatment.

The So What?

So what does an Australian critical infrastructure operator actually do on Monday? The Enhanced Critical Infrastructure Risk Management Program Rules made this year already ask for most of it in prose, across nine asset classes that now explicitly include liquid fuel, with grace periods of 12 and 24 months already ticking (Department of Home Affairs, 2026). The regulator is asking for the same graph.

  • Declare the graph. Map your supply chain for major suppliers and critical components, exactly as the s 10A supply-chain hazard obligation now requires (Department of Home Affairs, 2026). Anything you leave off the graph is attack surface you will meet by surprise.

  • Verify the edges. Assess each supplier's access, influence and control, including foreign ownership, control or influence, and back it with signed provenance where the edge is software (Department of Home Affairs, 2026). A declared edge you have not verified is a rumour.

  • Make the build hermetic. Turn declaration into enforcement so the build cannot reach a dependency you did not list. This is the move that makes the graph honest (Open Source Security Foundation, 2023).

  • Measure days of cover for digital dependencies. You know your diesel number. What is your 34 days for a critical software-as-a-service provider, a model API, a certificate authority? If you cannot state it, you have not finished declaring the graph.

  • Threat-model the opportunist. Harden your weakest seam before your scariest adversary, because the criminal triggered by the crisis will get there first (European Union Agency for Cybersecurity, 2025). Pair it with Essential Eight Maturity Level Two and phishing-resistant multi-factor authentication, which the rules now mandate (Department of Home Affairs, 2026).

  • Assure continuously, not annually. Treat the graph as living. Monitor provenance the way you monitor stock levels, not the way you file a return.

  • Point AI at the graph, not at the gap. Use the model for entity resolution, edge extraction, anomaly detection and simulation over what you have declared. Never as a substitute for declaring it.

Want to feel the argument rather than read it? The interactive below lets you fail a node in a live dependency graph and watch what goes dark downstream. Try knocking out an edge you never knew was load-bearing.

Interactive: fail a node, close the strait, then declare the graph and keep it verified, and watch what the monitor could see all along.

Conclusion

The fuel crisis was the rare emergency where this country could argue from a position of knowledge. When the strait closed, we had a number, because unglamorous people did unglamorous declaration work decades ago and kept it continuously queryable. That is not luck. That is a control, quietly doing its job for forty years.

Your software estate, your defence supply chain, your cloud build, gets no such gift by default. The graph is there whether or not you have drawn it. The dependencies are real whether or not you declared them. The only variable is whether you find out on a quiet Tuesday, by running a query, or on the worst morning of the year, by drawing the map while the building burns.

So here is the question I would leave you with. When your strait closes, and one of them will, will your first day be spent querying a graph, or drawing one?

Thanks for reading.

The views expressed in this article are my own and do not represent those of my employer or any of my clients.

References

Related reading

9 min read

Exempt by Design: The AI Governance Gap in the National Intelligence Community

The TL;DR is that we assume the most consequential users of AI must be the most tightly governed, and in Australia it runs the other way. The mandatory framework for government AI, the Digital Transformation Agency's policy now at v2.0 with a mandatory use-case register, exempts the Defence portfolio and the National Intelligence Community (NIC) (Digital Transformation Agency, 2025). The National AI Plan shelved the proposed mandatory high-risk guardrails and stood up an AI Safety Institute instead (Department of Industry, Science and Resources, 2025). What is left for the community is point-in-time authorisation under the Protective Security Policy Framework and an oversight office that has appointed a single Chief AI Officer (Inspector-General of Intelligence and Security, 2025).

AIGovernanceNational SecuritySovereign AIOversight
7 min read

The 80,000-Entity Wave: Why Tranche 2 Is an Entity Resolution Problem, Not a Compliance One

The TL;DR is that the anti-money-laundering reforms known as Tranche 2 are being sold to the newly captured professions as an onboarding and compliance exercise, when the real problem sits downstream, inside the financial intelligence system that now has to resolve, monitor and make sense of roughly 80,000 new reporting entities without drowning its analysts in false positives (Australian Transaction Reports and Analysis Centre, 2025).

AIEntity ResolutionAML/CTFGraph
1 min read

The Double-Edged Sword of AI in Law Enforcement: Navigating the Deep Fake Dilemma

The same AI that helps investigators can also fabricate convincing evidence. A look at how deepfakes cut both ways in law enforcement, and what accountable governance requires.

National SecurityDeepfakesLaw Enforcement