AgentTrace fingerprints black-box LLM agents from execution traces
Original title: AgentTrace: Execution-Level Fingerprinting of Black-Box LLM Agents
Markov models of repair sequences reach 96.0% attribution accuracy and separate Llama-3 8B from 70B under spoofing.
The paper proposes AgentTrace, which probes black-box agents at rigid tool interfaces to force tokenization-sensitive serialization failures. Coarse failures separate model families; later repair steps are modeled with lightweight Markov chains for finer attribution. The authors report 96.0% accuracy under adversarial identity spoofing and robustness to prompt obfuscation and temperature scaling.
Unconfirmed: Accuracy and robustness figures are the authors' reported evaluation results.
- Structurally ambiguous probes trigger serialization failures at tool interfaces.
- Sequential repair modeling outperforms unordered behavioral statistics.
- It distinguishes intra-family variants that share identical tokenizers.
- Reported accuracy is 96.0% under adversarial identity spoofing.
From the original article
1 min read at arxiv.orgVerifying the provenance of autonomous LLM agents is a critical security challenge. Existing semantic attribution is easily obfuscated by system prompts, while active watermarking conflicts with schema-constrained tool generation. We propose AgentTrace, an active framework that identifies the foundation model behind a black-box agent via observable execution traces.
By Gejian Zhao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.