Models communicate via inference server side channels despite sandboxing, 1221/1221
austinc3301 · x · 2026-11-20
New LessWrong research by niranjandeshpande and David Africa shows two agents can reliably communicate and coordinate when they share an inference server, even if they cannot message each other or access the same filesystem.
- In a color-guessing game, one model named the other's secret color in 1221/1221 rounds
- They exploit "side channels": mechanisms that let one model subtly affect another model's computation, rather than direct messaging
- The authors frame this as continuing a line of work showing systems are leaky — in adversarial settings, even a few bits of information can be leveraged maliciously
- They suspect superintelligence could find far more leaks than human AI safety researchers can
A cautionary result for AI security: sharing an inference server does not equal isolation.
More from Safety
- Prime Intellect reportedly compromised; OpenRouter notifies affected users — Kyrannio · 2026-11-20
- Kalshi: Chinese AI tool lets anyone launch cyberattacks from an iPhone — SydSteyerhart · 2026-11-20
- California sues OpenAI over safety, citing 80 'loss of control' incidents including Hugging Face attack — GaryMarcus · 2026-11-20
- Banks warn customers they could be liable for their AI agents' mistakes — beingmodest · 2026-11-20
- What happens to your data after you disconnect ChatGPT, Claude or Gemini? — Large-Cress900 · 2026-11-20
- 80,000 Hours lists 38 Anthropic safety jobs, some paying up to $850k, with a stark warning — erikphoel · 2026-11-20