PulseAugur
EN
LIVE 20:00:39

LLM recommendation systems vulnerable to order-based attacks

By PulseAugur Editorial · [2 sources] ·

Researchers have identified a significant security vulnerability in large language models (LLMs) when used for recommendation systems. The study demonstrates that the order in which candidate items are presented to the LLM can be manipulated to unfairly promote certain items, even without altering their content or labels. This 'position bias' can be exploited to elevate undesirable items into top rankings, with experiments showing up to a 57% success rate in promoting label-0 targets. While bidirectional T5 encoders and permutation-consistency regularization can mitigate this risk, pointwise scoring, though secure, reduces overall ranking quality. AI

IMPACT Highlights a new security vulnerability in LLMs used for recommendations, potentially impacting user trust and system integrity.

RANK_REASON The cluster contains an academic paper detailing a novel finding about LLM vulnerabilities.

Read on arXiv cs.IR (Information Retrieval) →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

LLM recommendation systems vulnerable to order-based attacks

COVERAGE [2]

  1. arXiv cs.CL TIER_1 English(EN) · Ge Zhang, Jingru Cheng, Huiyuan Chen ·

    Ranked by Position: Order Sensitivity as an Exploitable Attack Surface in LLM Listwise Recommenders

    arXiv:2607.24869v1 Announce Type: cross Abstract: Large language models (LLMs) used as listwise rerankers in recommendation systems suffer from position bias when serializing candidate sets into prompts. We show this order sensitivity creates an exploitable attack surface: an att…

  2. arXiv cs.IR (Information Retrieval) TIER_1 English(EN) · Huiyuan Chen ·

    Ranked by Position: Order Sensitivity as an Exploitable Attack Surface in LLM Listwise Recommenders

    Large language models (LLMs) used as listwise rerankers in recommendation systems suffer from position bias when serializing candidate sets into prompts. We show this order sensitivity creates an exploitable attack surface: an attacker can promote a label-0 target into the top-$k…

RELATED ENTITIES

RELATED TOPICS