Supply chain risk
Also called Supply Chain Risk
The danger that something you didn’t build — a dataset, a base model, a dependency — arrives already compromised.
Think of it like
Your kitchen can be spotless, but if a supplier ships tainted ingredients, the meal is still poisoned.
Example
A team fine-tunes a popular open-weights model that, unknown to them, carries a trojan planted upstream.
How it actually works
Modern AI systems inherit trust from pretrained weights, scraped datasets, libraries, and model hubs — any of which an attacker can tamper with (poisoned data, backdoors, malicious packages). Because the compromise happens before you touch it, your own testing may not catch it. Provenance, checksums, and vetting sources are the defenses.
For product teams
You inherit the security of everything you didn’t build yourself — including the base model.
For engineers
Pretrained weights, datasets, and dependencies are trust-inherited attack vectors; verify provenance and integrity, not just behavior.
Related
- Poisoned data — Includes ingesting poisoned data.
- Backdoor attack — Can deliver a backdoor attack.
- Model provenance — Countered by tracking model provenance.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome