The Trigger in the Haystack: Extracting and Reconstructing LLM Backdoor Triggers
This addresses AI security by providing a scalable detection method for poisoned models, though it is incremental as it builds on existing memory extraction and pattern analysis techniques.
The paper tackles the problem of detecting sleeper agent-style backdoors in causal language models by developing a practical scanner that identifies backdoor triggers without prior knowledge, showing it recovers working triggers across various scenarios and models.
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.