Reasoning Introduces New Poisoning Attacks Yet Makes Them More Complicated

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Foerster, Hanna, Shumailov, Ilia, Zhao, Yiren, Chaudhari, Harsh, Hayes, Jamie, Mullins, Robert, Gal, Yarin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914026110320640
author Foerster, Hanna
Shumailov, Ilia
Zhao, Yiren
Chaudhari, Harsh
Hayes, Jamie
Mullins, Robert
Gal, Yarin
author_facet Foerster, Hanna
Shumailov, Ilia
Zhao, Yiren
Chaudhari, Harsh
Hayes, Jamie
Mullins, Robert
Gal, Yarin
contents Early research into data poisoning attacks against Large Language Models (LLMs) demonstrated the ease with which backdoors could be injected. More recent LLMs add step-by-step reasoning, expanding the attack surface to include the intermediate chain-of-thought (CoT) and its inherent trait of decomposing problems into subproblems. Using these vectors for more stealthy poisoning, we introduce ``decomposed reasoning poison'', in which the attacker modifies only the reasoning path, leaving prompts and final answers clean, and splits the trigger across multiple, individually harmless components. Fascinatingly, while it remains possible to inject these decomposed poisons, reliably activating them to change final answers (rather than just the CoT) is surprisingly difficult. This difficulty arises because the models can often recover from backdoors that are activated within their thought processes. Ultimately, it appears that an emergent form of backdoor robustness is originating from the reasoning capabilities of these advanced LLMs, as well as from the architectural separation between reasoning and final answer generation.
format Preprint
id arxiv_https___arxiv_org_abs_2509_05739
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Reasoning Introduces New Poisoning Attacks Yet Makes Them More Complicated
Foerster, Hanna
Shumailov, Ilia
Zhao, Yiren
Chaudhari, Harsh
Hayes, Jamie
Mullins, Robert
Gal, Yarin
Cryptography and Security
Artificial Intelligence
Machine Learning
Early research into data poisoning attacks against Large Language Models (LLMs) demonstrated the ease with which backdoors could be injected. More recent LLMs add step-by-step reasoning, expanding the attack surface to include the intermediate chain-of-thought (CoT) and its inherent trait of decomposing problems into subproblems. Using these vectors for more stealthy poisoning, we introduce ``decomposed reasoning poison'', in which the attacker modifies only the reasoning path, leaving prompts and final answers clean, and splits the trigger across multiple, individually harmless components. Fascinatingly, while it remains possible to inject these decomposed poisons, reliably activating them to change final answers (rather than just the CoT) is surprisingly difficult. This difficulty arises because the models can often recover from backdoors that are activated within their thought processes. Ultimately, it appears that an emergent form of backdoor robustness is originating from the reasoning capabilities of these advanced LLMs, as well as from the architectural separation between reasoning and final answer generation.
title Reasoning Introduces New Poisoning Attacks Yet Makes Them More Complicated
topic Cryptography and Security
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2509.05739