![The AI Supply Chain Is Getting Weird [Monthly Guest Post]](https://substackcdn.com/image/fetch/$s_!XVf6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe199e306-ab5d-4239-938e-59f5535a5497_1024x301.jpeg "The AI Supply Chain Is Getting Weird [Monthly Guest Post]")
Data poisoning targets AI systems during training by corrupting datasets or models, allowing attackers to embed backdoors, mislabeled data, or hidden behaviors that can survive deployment and evade standard benchmarks. The article explains how public model repositories, third-party datasets, fine-tuning pipelines, and RAG sources expand the AI supply chain risk, and highlights controls such as provenance checks, red-teaming, and an AIBOM. #HuggingFace #JFrog #PoisonGPT #Mercor #OWASPLLMTop10 #AIBOM
Keypoints
- Data poisoning alters AI training data before deployment.
- Attackers can plant backdoors, wrong labels, or subtle manipulation.
- Only a small amount of poisoned data may be enough to affect a model.
- Public model repositories and third-party datasets increase supply chain risk.
- Provenance checks, red-teaming, and AIBOMs help reduce exposure.
Read More: https://www.toxsec.com/p/the-ai-supply-chain-is-getting-weird