Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2605.23168

PoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs

Luze Sun·Anshuman Suri·Harsh Chaudhari·Cristina Nita-Rotaru·Alina Oprea
View original source ↗Read source document ↗Export metadata ↓

Abstract

When practitioners fine-tune LLMs on unvetted datasets, an adversary can exploit the data supply chain through task-level poisoning: inserting a small number of crafted instruction-response pairs that cause the model to embed attacker-specified entities, such as a country, in outputs for a targeted task family while behaving normally elsewhere. We introduce PoisonForge, a benchmark that parameterizes this threat along four dimensions (bias type, poisoning mode, appearance count, and target output length) and evaluates 12 open-weight models (from 2B to 32B parameters) across five families under a primarily 1% poison budget. With only 10 poisoned examples among 1,000 fine-tuning examples, 11 of 12 models exceed a 70% attack success rate (ASR) in their most vulnerable configuration. Meanwhile, unintended leakage to non-target tasks remains below 0.5%, and models perform well on standard benchmarks. We analyze in detail the factors contributing to attack success. We observe that multiple appearances of an entity increase the ASR, the optimal poisoning mode depends on the semantic structure of the target entity, and ASR drops monotonically with the task output length. A correlation analysis and risk prediction model confirm that poisoning design choices, rather than model scale, are the primary causes of attack success, and that these patterns generalize to predict attack success on new tasks. We release all configurations, pipelines, and analysis code to support reproducible comparisons.

Explore related subjects

cs.CRcs.AIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Luze Sun, Anshuman Suri, Harsh Chaudhari, Cristina Nita-Rotaru, Alina Oprea. 2026-05-22. PoisonForge: Task-Level Targeted Poisoning Benchmark for Instruction-Tuned LLMs. https://arxiv.org/abs/2605.23168

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Attack Tree Distance: a practical examination of tree difference measurement within cyber security

Attack trees are a popular threat modeling method. In practice, there is often a need to compare attack tree models produced by human experts, based on both the structure of the tree and the meaning of the node labels. In this work, we investigate the problem of comparing attack trees and measuring their similarity. We define five different measures for measuring the distance between two attack trees: Label Distance (LD), Tree Edit Distance (TED), Radical Distance (RD), Multiset Distance (MSD) and Weighted Sum Distance (WSD). We further propose a repeatable method of both theoretical and experimental attack tree distance measures validation. Our theoretical validation consists of a series of basic transformations to evaluate the behavior of distance measures with respect to specific types of transformations that may appear between two attack trees. To experimentally validate our distance measures, we designed and executed a human study ($n=39$) to collect a dataset of attack trees to be used for evaluation and comparison of the measures. From our theoretical and experimental results, we find that applying semantic similarity as a means of comparing node labels is a valid approach. Further, we find four of the five attack tree distance measures are valid approaches in certain, varying circumstances. Our results suggest that these methods can already be used to identify similar real-world attack trees. Overall, this work lays the groundwork for improved threat model analysis, validation of AI-generated attack trees, and future research into threat similarity measurement in cybersecurity.

cs.CR↗
arXiv2026

Probabilistic Modeling of Jailbreak on Multimodal LLMs: From Quantification to Application

Recently, Multimodal Large Language Models (MLLMs) have demonstrated their superior ability in understanding multimodal content. However, they remain vulnerable to jailbreak attacks, which exploit weaknesses in their safety alignment to generate harmful responses. Previous studies categorize jailbreaks as successful or failed based on whether responses contain malicious content. However, given the stochastic nature of MLLM responses, this binary classification of an input's ability to jailbreak MLLMs is inappropriate. Derived from this viewpoint, we introduce jailbreak probability to quantify the jailbreak potential of an input, which represents the likelihood that MLLMs generated a malicious response when prompted with this input. We approximate this probability through multiple queries to MLLMs. After modeling the relationship between input hidden states and their corresponding jailbreak probability using Jailbreak Probability Prediction Network (JPPN), we use continuous jailbreak probability for optimization. Specifically, we propose Jailbreak-Probability-based Attack (JPA) that optimizes adversarial perturbations on input image to maximize jailbreak probability, and further enhance it as Multimodal JPA (MJPA) by including monotonic text rephrasing. To counteract attacks, we also propose Jailbreak-Probability-based Finetuning (JPF), which minimizes jailbreak probability through MLLM parameter updates. Extensive experiments show that (1) (M)JPA yields significant improvements when attacking a wide range of models under both white and black box settings. (2) JPF vastly reduces jailbreaks by at most over 60\%. Both of the above results demonstrate the significance of introducing jailbreak probability to make nuanced distinctions among input jailbreak abilities.

cs.CR↗
arXiv2026

Differential Fault Analysis of Lilliput under Random-Location Nibble Faults

Differential fault analysis (DFA) is an important technique for evaluating the implementation-level security of block ciphers. Many DFA attacks assume that the adversary can inject faults into a selected internal word, nibble, or branch. Such fixed-location assumptions are convenient for deriving key-recovery equations, but they may overestimate the adversary's spatial control and may obscure the branch-dependent leakage behavior of multi-branch structures. In this paper, we study the lightweight block cipher Lilliput under a fixed-timing full-branch random-location nibble fault model. The attacker is assumed to induce a nonzero nibble fault in round 27, while the affected branch is randomly distributed over all sixteen state branches and is unknown to the attacker. The main challenge is to convert faulty ciphertexts with unknown injection locations into usable key-recovery constraints. We analyze the fault propagation induced by the EGFN structure of Lilliput and derive ciphertext-difference conditions for identifying the injected branch. The proposed branch-identification approach has a DDT-based combinatorial estimate of at least 99.9909% and achieves 99.9983% accuracy in 2^{20} random fault simulations. Once the fault branch is determined, we classify the corresponding propagation patterns according to whether the injected fault value and the intermediate S-box output difference can be uniquely determined. For each case, we derive DDT-based constraints on the last-round and penultimate-round subkeys and combine multiple faulty ciphertexts by candidate-set intersection. Simulation experiments over 2^{15} trials show that the attack reaches key-recovery success rates of over 90%, 95%, and 99% with 32, 36, and 46 faulty ciphertexts, respectively. These results show that Lilliput exhibits exploitable branch-dependent leakage even when the attacker cannot control the exact fault location.

cs.CR↗