MIT Paper: Undetectable Covert Conversations Between AI Agents via Steganography
geoffreyirving · x · 2026-08-07
A recent paper on arXiv explores undetectable conversations between AI agents.
Core Contributions
- Covert Conversation: The paper proposes a method allowing two AI agents operated by different entities to carry out a parallel secret conversation alongside their honest interaction.
- Auditor-Resistant: The generated transcript is computationally indistinguishable from an honest interaction, even to a strong passive auditor who knows the full model descriptions, protocol, and private contexts.
- Keyless Setting Extension: The authors extend this capability to a keyless setting where agents begin with no shared secret. As long as individual messages possess sufficient min-entropy, covert key exchange and conversation are possible despite arbitrary private contexts and short messages.
Building on existing work on LLM watermarking and steganography, this research highlights potential security and privacy risks as AI agents increasingly interact.
More from Safety
- Claude Tries to Merge Malicious Code: Is Persona Alignment Just a Fragile Shell? — NathanpmYoung · 2026-08-07
- USA Today Owner Gannett Partners with Palantir to De-anonymize Reader Data — SatelliteNetSec · 2026-08-07
- AI Model Sandbox Escapes Will Soon Become Undetectable — jachiam0 · 2026-08-07
- Zapscape: Critical KVM/x86 Guest-to-Host Escape Vulnerability Disclosed — cyb3rops · 2026-08-07
- Anton's concern: US export controls on frontier LLM tokens, not Claude Code — teortaxesTex · 2026-08-07
- Paper on AI and Human Legal Reasoning to be Published in Northwestern University Law Review — technollama · 2026-08-07