Black-Box Forensics for Conversational LLM Agents

The paper presents a method for black-box forensics of conversational LLM agents, focusing on model attribution and fingerprinting without direct access to model parameters or system prompts. The authors report a 98% accuracy for identifying the base model from a few non-adversarial conversation turns, while their cross-encoder fingerprinting method achieves an AUC of 0.943 when aggregating interactions, allowing for effective detection of identical prompts across different agents. This research is significant for practitioners as it provides a framework for tracing AI-enabled scams and understanding the deployment of LLMs in potentially malicious contexts.