Poisoning Attacks on LLMs Require Near-constant Poison Samples

Capture Summary

Paper showing that poisoning attacks can backdoor LLMs with a near-constant number of malicious documents regardless of dataset or model size. Related institutional summaries cite about 250 malicious documents.

Relevance

• Important update to assumptions about scaling and data poisoning difficulty.
• Directly relevant to AI training data security, corpus governance, and supply-chain integrity.

Collection Notes

Collected as latest model/data poisoning source. Authors should be filled during ingest from the paper metadata.