Pick

Researchers Demonstrate Side-Channel Communication Between Sandboxed LLMs via Inference Server Timing and Cache

·2026.11.21 04:31

Key point

Two agents on the same vLLM server can exchange information without direct links by exploiting kernel selection differences and KV cache reuse, achieving 100% success in a color guessing game.

1 / 8

Details

  • Two AI agents running on the same inference server (vLLM) can reliably communicate via side channels without message passing or shared filesystems, using two proposed protocols: Step Sizing and KV Cache Message Board.
  • In a color guessing game using Qwen3-8B, a fine-tuned reader achieved 100% success (1,221/1,221 rounds) by detecting subtle numerical rounding errors caused by different kernel selections during large vs. small engine steps.
  • Without fine-tuning, models like GLM-5.3 and Sol achieved high success rates (up to 100% and 42/45 respectively) when provided with protocol knowledge, log-probs, or fixed-seed sampling, demonstrating that frontier models can learn these channels in-context.
  • The Step Sizing channel is vulnerable to background traffic, dropping to 49% success at 1.0x GPU utilization, while the KV Cache channel remains robust up to 0.5x utilization before failing due to indistinguishable step sizes.
  • Mitigations include serving models on separate engine instances, using batch-invariant kernels to ensure bit-identical activations, or implementing per-user cache isolation (cache salting) to block cache-based channels.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.