Persistent Pre-training Poisoning of LLMs

Yiming Zhang, Javier Rando, Ivan Evtimov, Jianfeng Chi, Eric Michael Smith, Nicholas Carlini, Florian Tramèr, Daphne Ippolito

Keywords: poisoning, pretraining, large language models, security

TL;DR: We poison LLMs during pretraining and show that an attacker with control of only 0.1% of the data can manipulate model behavior even after alignment.

Abstract:

Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web.
Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets.
Our work evaluates for the first time whether language models can also be \emph{compromised during pre-training}, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO).
We pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B).
Our main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10398