AI/ ai · coding-agents · security · llm-fingerprinting

Researchers Fingerprint LLMs by How They Behave, Not What They Say

A new method identifies which AI model powers a coding agent by watching its behavior, not its words.

Turns out you can tell which AI model is driving a coding agent just by watching how it works, not what it writes.

Researchers built a system called LIDAR that fingerprints large language models running inside coding-agent harnesses - the tools that let an LLM inspect a repository, call functions, and edit files. Instead of analyzing a model's text output or token probabilities, LIDAR runs three paired probes that trigger controlled scenarios: verifying a change after an edit, recovering from a transient failure, and resolving conflicts between a specification and a test. It compares the resulting trajectories against clean reference behavior using a lightweight probabilistic classifier, and needs no access to model weights, logits, or provider internals - it works purely by observing black-box behavior. Tested across 36 models from seven families and two separate agent harnesses, LIDAR outperformed four existing fingerprinting and API-auditing baselines on accuracy.

Swapping the model behind an agent can quietly change how it handles risk - whether it verifies its own edits or recovers safely from a failure - and most users have no way to know which model is actually doing the work. That is a real problem for anyone trusting a tool marketed as running on a specific model without a way to check the claim, and it hands auditors a method for catching providers who quietly substitute cheaper models behind the scenes.

It is a clever trick, though the same technique that catches a mislabeled agent could just as easily help a rival fingerprint - and copy - your stack.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →