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

A research paper demonstrates practical attacks that fingerprint inference engines and exploit model-driven environmental discovery, potentially enabling escape from sandboxed AI systems.

Auto-scored: Demonstrates a real escalation in AI security by showing practical fingerprinting and escape techniques, though not yet weaponized or causing documented harm.