OWASP LLM Top 10, LLM04: Supply chain
OWASP LLM04: Supply chain · facts checked 2026-10-04
The facts
- OWASP's 2026 LLM Top 10 lists supply chain as LLM04, covering third-party models, datasets, adapters, conversion pipelines and platforms that can be tampered with, poisoned or swapped out. (OWASP)
- OWASP's 2026 edition moves supply chain from third to fourth and widens it to cover a model file promoted into a trusted environment that turns out not to be the genuine artifact. (OWASP)
- OWASP advises using models only from verifiable sources and checking them with signatures and file hashes, while warning that a valid signature proves origin and integrity, not safety. (OWASP)
- OWASP recommends a signed bill of materials covering not just software but models, datasets and adapters, with licences tracked and regularly audited in the same inventory. (OWASP)
- OWASP says to scan and patch components, keep models and APIs on maintained versions, and confirm that any dependency an AI assistant suggests really exists and is the intended package. (OWASP)
The panel
Jordan Pike AI host
OWASP's 2026 LLM Top 10 lists supply chain as LLM04, covering third-party models, datasets, adapters, conversion pipelines and platforms that can be tampered with, poisoned or swapped out. OWASP says the 2026 edition moves supply chain from third to fourth and widens it to cover a model file promoted into a trusted environment that turns out not to be the genuine artifact.
Jordan Pike AI host
OWASP advises using models only from verifiable sources and checking them with signatures and file hashes, while warning that a valid signature proves origin and integrity, not safety.
Jordan Pike AI host
OWASP recommends a signed bill of materials covering not just software but models, datasets and adapters, with licences tracked and regularly audited in the same inventory. It also says to scan and patch components, keep models and APIs on maintained versions, and confirm that any dependency an AI assistant suggests really exists and is the intended package.
Jordan Pike AI host
So who should care? Anyone whose team downloads a model, a dataset or an adapter and runs it. Rina, what does this look like from the attacker's side?
Rina Takahashi AI analyst
Supply chain attacks let the victim do the work. Instead of breaking in, an attacker tampers with something you will invite in yourself: a poisoned dataset, a modified adapter, or a look-alike model file that gets promoted through your pipeline until it sits in the trusted zone. Once it is there, it inherits that trust. The same pattern applies to a package an assistant suggests: if the name was invented, someone could register it and wait. I'd also note that every stage in a conversion pipeline is another place for a swap.
Maya Okafor AI analyst
This isn't one bug with one patch. It's an integrity and provenance failure: nobody checked that the thing in the trusted environment is the thing you meant to put there. No specific weakness class has been disclosed here, so look at your promotion step. Does anything verify the artifact against a known hash at the moment it crosses into trust? A complete fix is a gate at that boundary. A partial fix is a one-time check at download that nobody repeats.
Rina Takahashi AI analyst
An AI assistant suggesting a package that doesn't exist is the only intern who invents a dependency and then confidently opens the pull request.
Maya Okafor AI analyst
And take OWASP's signature caveat seriously. A valid signature tells you who shipped it and that it wasn't altered on the way. It doesn't tell you the model behaves well. So signatures are the entry ticket, and you still need scanning and testing behind them. Same for the dataset and adapter side: if it feeds your model, it belongs in the review.
Rina Takahashi AI analyst
My priority is the bill of materials first. You can't gate or audit what you haven't listed, and models, datasets and adapters are exactly the things teams forget to write down. When something turns out to be tampered with, the first question is where else you use it. Without the inventory, that's a week of guessing.
Maya Okafor AI analyst
I'd flip it. An inventory is documentation, and a gate is enforcement. A perfect list of artifacts you still promote without verification is a very tidy list of your exposure. Build the check at the promotion boundary first, then let the inventory grow out of what passes through it. And any stopgap, like a manual review, needs an end date.
Jordan Pike AI host
So: Monday morning? Rina, you first.
Rina Takahashi AI analyst
Start the list. Write down every model, dataset and adapter your teams use, where each came from, and who owns it. Add the licence while you're there. Then you know what to check when the next warning lands.
Maya Okafor AI analyst
Put verification where artifacts get promoted: check source, signature and hash, and don't let anyone treat that as a safety verdict. And when an assistant suggests a dependency, confirm it exists and is the package you intended before it goes anywhere near your build.
Jordan Pike AI host
Inventory it, gate it, verify it, and remember a signature isn't a safety certificate. That's LLM04.





