Q-Cyber Magazine
The AI supply chain is now a security inventory problem
OWASP's 2025 LLM Top 10 and the 2026 SBOM minimum elements, which flag AI systems as needing additional fields, give organisations a public basis for inventorying models, data, tools and dependencies.

An AI application is rarely a single artefact. It is a model, a set of weights, training or retrieval data, a toolchain, and a cluster of ordinary software dependencies. If those ingredients are unnamed, their risk cannot be managed.
Two lists, one operational habit
OWASP's list is a risk taxonomy. It reminds builders that models, datasets, plugins and service providers can be a source of weakness, and that pre-training, fine-tuning or embedding data can be poisoned so that the model behaves in ways the deployer did not intend. The 2026 SBOM minimum elements are a data standard. They ask producers and buyers to record, at least, the author and signature of the inventory, and, for each component, the producer, name, version, identifiers, dependencies, hashes and licence, with unknowns stated explicitly.
Read together, they recommend a habit: do not run a model whose origin, version and dependencies you cannot name. For AI, the inventory should extend beyond libraries to weights, checkpoints, tokenisers, retrieval corpora and fine-tuning sets, which is precisely why the 2026 guidance says AI systems may need additional fields, while insisting that work should still begin with the minimum elements.
CERT-In's reference to AIBOM, in a public national framework, is further evidence that this is no longer only a foreign procurement idea. Indian organisations can already ask suppliers for a software bill of materials that meets the 2026 minimum, and for an AI-specific inventory where a model is part of the system.
Exhibit 1
What an AI inventory should be able to answer
| Question | Public reference | Typical artefact |
|---|---|---|
| Which software components are present? | 2026 SBOM minimum elements (all software) | Signed SBOM with producer, version, dependencies, hashes and licence |
| Where might AI need more than a software list? | 2026 SBOM guidance: AI systems may require additional elements | Extended inventory of models, weights, data sets and tools |
| What can go wrong in that chain? | OWASP LLM03 (supply chain) and LLM04 (data and model poisoning) | Named review of model source, training or fine-tuning data, and third-party tools |
| Is there an Indian public hook? | CERT-In technical guidelines on SBOM and AIBOM, cited in the February 2026 space framework | Contract clause requiring those artefacts at delivery and at each material update |
Source: CISA and partners, 2026 Minimum Elements for a Software Bill of Materials, 29 July 2026; OWASP Gen AI Security Project, 2025 LLM Top 10; CERT-In space and SatCom framework, February 2026, section 5.8.
Implications
- Start with the 2026 SBOM minimum elements, then add model- and data-specific fields.
- Name the source of weights and fine-tuning data, and treat "unknown" as a risk to be accepted explicitly.
- Review poisoning and third-party tools as part of change control, not only as a research topic.
- Refresh the inventory when the model, the retrieval corpus or the toolchain changes.
What this means for AI
An organisation that cannot list the models and data it runs is not yet in a position to claim control of them. The public standards needed to start that list already exist. The remaining work is to put them in contracts and in operations.


