AgentTrace preprint reports 96.0% accuracy in identifying agent models

Researchers report that AgentTrace identified the foundation model behind a black-box AI agent with 96.0% attribution accuracy under adversarial identity spoofing. The preprint describes a method that examines observable execution traces rather than relying on an agent's stated identity or answer wording. AgentTrace sends structurally ambiguous probes that trigger formatting failures at strict tool interfaces, then models the sequence of repair attempts with lightweight Markov chains. Those recovery patterns distinguished evaluated variants sharing identical tokenizers, including Llama-3 8B/70B. The researchers report robustness against prompt obfuscation and temperature scaling; the findings apply to the evaluated models and probing setup.