Analysis
A small number of samples can poison LLMs
First reported · Discovered anthropic.com
Page published
Earliest dated coverage: 6 Oct 2026 · First observed: 16 Oct 2026 · Latest dated coverage: 6 Oct 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
AI/ML supply-chain defenses must address both executable artifacts and manipulated model behavior, since conventional dependency and artifact inspection cannot establish that learned behavior is trustworthy.
Supply Chain Poisoning in AI/ML Pipelines surveys risks across training data, dependencies, and pretrained models, distinguishing code execution during artifact loading from poisoned model behavior. The overview cites research showing that 250 malicious pretraining documents could implant a narrow gibberish-output backdoor in models ranging from 600M to 13B parameters.