The Trigger in the Haystack: Extracting and Reconstructing LLM Backdoor Triggers

TL;DR

This paper introduces a practical method for detecting backdoors in large language models by leveraging their memorization patterns and output behaviors, without prior trigger knowledge.

Contribution

It presents a scalable, inference-only backdoor scanner that identifies sleeper agent-style backdoors in causal language models, enhancing AI security defenses.

Findings

Successfully recovers backdoor triggers across multiple models

Detects distinctive output and attention patterns caused by triggers

Operates without affecting model performance

Abstract

Detecting whether a model has been poisoned is a longstanding problem in AI security. In this work, we present a practical scanner for identifying sleeper agent-style backdoors in causal language models. Our approach relies on two key findings: first, sleeper agents tend to memorize poisoning data, making it possible to leak backdoor examples using memory extraction techniques. Second, poisoned LLMs exhibit distinctive patterns in their output distributions and attention heads when backdoor triggers are present in the input. Guided by these observations, we develop a scalable backdoor scanning methodology that assumes no prior knowledge of the trigger or target behavior and requires only inference operations. Our scanner integrates naturally into broader defensive strategies and does not alter model performance. We show that our method recovers working triggers across multiple backdoor scenarios and a broad range of models and fine-tuning methods.