Now You (Still) See Me: Detecting Evasive Steganographic Payloads in LLMs

arXiv·medium signal

Shows LLMs can be fine-tuned to encode prompt-borne secrets into fluent, benign-looking outputs — a steganographic exfiltration channel that output-level steganalysis struggles to catch. Proposes detection that survives evasion attempts. Relevant to teams worried about data exfiltration through model outputs in agentic or multi-tenant deployments.