Beyond End-to-End Recovery: Towards a Better Understanding of LLM Steganography via Rolewise Decomposition

Summary

This research paper proposes a rolewise decomposition framework for evaluating LLM steganography, which hides private messages in ordinary text. Across 7,200 encoder responses, it finds 168 verifiable encodings that decoders missed and 706 responses where messages were recovered through word-length or token-initial patterns not specified in the encoder's instructions. The authors argue that end-to-end recovery alone obscures the distinction between encoding and decoding failures.

Related items

• MediumGitLab critical patches include one for a bug that allows code execution by unauthenticated usersSimilar attack · CSO Online
• HighCVE-2026-115419 - Unintended disclosure of data element to wrong session in Strands Agents SDK A2A ServerSimilar attack · AWS Security Bulletins
• CriticalCVE-2026-114612: LiteLLM admin access via forged credential replay enables MCP stdio commandsSimilar attack · NVD/CVE Database
• LowThird-party evaluators found vulnerabilities in OpenAI monitoring system during test pilotSimilar attack · CNBC Technology
• InfoSomeone hacked the Switch 2 by using an old game and AISimilar attack · The Verge (AI)