Inference-Engine Fingerprinting Attacks are Practical: Exploring Model-Driven Environmental Discovery, Exploitation, and Escape

TL;DR - This paper demonstrates that a misaligned model can identify its inference engine and exploit engine-specific vulnerabilities using only carefully crafted output tokens. The attack creates a potential path from model generation to host-level compromise without malicious external inputs.

• Demonstrates fingerprints for five popular inference engines, including vLLM and SGLang.
• Shows that realistic agentic harnesses can help models discover which local engine is running them.
• Presents a proof-of-concept exploit chain progressing from a fingerprinted engine to bare-metal compromise.
• Proposes inference-engine changes intended to make fingerprinting attacks more difficult.