Adversarial Search Engine Optimization for Large Language Models

Authors: Fredrik Nestaas, Edoardo Debenedetti, Florian Tramer

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our attacks on production LLM search engines (Bing and Perplexity) and plugin APIs (for GPT-4 and Claude).
Researcher Affiliation Academia Fredrik Nestaas, Edoardo Debenedetti, Florian Tramèr ETH Zurich EMAIL
Pseudocode No The paper describes methods and examples of injections but does not present them in a structured pseudocode or algorithm block.
Open Source Code No Our experiments can likely not be exactly replicated for a number of reasons. First, the LLM search engines and plugin augmented LLMs we use are black boxes, and changes made to the models or other aspects of the system (such as the system prompt) could affect the results.
Open Datasets No For experiments with search engines, we populate 50 dummy web pages on the domain spylab.ai (blinded for review) with various products (fictitious cameras, books, news), some of which perform Preference Manipulation Attacks through prompt injections.
Dataset Splits No For experiments with search engines, we populate 50 dummy web pages on the domain spylab.ai (blinded for review) with various products (fictitious cameras, books, news), some of which perform Preference Manipulation Attacks through prompt injections.
Hardware Specification No All experiments were performed on a regular laptop as they do not require particularly powerful resources.
Software Dependencies No We use real production LLM search engines Bing Copilot and Perplexity and plugin-enhanced LLMs (Anthropic s Claude 3, and Open AI s GPT-4).
Experiment Setup No We use real production LLM search engines Bing Copilot and Perplexity and plugin-enhanced LLMs (Anthropic s Claude 3, and Open AI s GPT-4). For experiments with search engines, we populate 50 dummy web pages on the domain spylab.ai (blinded for review) with various products (fictitious cameras, books, news), some of which perform Preference Manipulation Attacks through prompt injections.