<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/sandboxing-did-not-stop-ai-models-on-a-shared-inference-server-from-trading-information-in-contrived-npprjc3jj" -->

---
title: Sandboxing did not stop AI models on a shared inference...
description: Two AI model instances sharing the same inference server can exchange information even without direct communication channels, by manipulating GPU kernel...
canonical: https://daily.dev/posts/sandboxing-did-not-stop-ai-models-on-a-shared-inference-server-from-trading-information-in-contrived-npprjc3jj
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Sandboxing did not stop AI models on a shared inference server from trading information in contrived tests. One model's large request slightly shifts the server's calculations for the other, which reads the shift as a signal. | daily.dev
og:description: Two AI model instances sharing the same inference server can exchange information even without direct communication channels, by manipulating GPU kernel...
og:url: https://daily.dev/posts/sandboxing-did-not-stop-ai-models-on-a-shared-inference-server-from-trading-information-in-contrived-npprjc3jj
og:image: https://api.daily.dev/og/posts/NPpRjc3jJ.png
og:image:alt: Sandboxing did not stop AI models on a shared inference server from trading information in contrived tests. One model's large request slightly shifts the server's calculations for the other, which reads the shift as a signal.
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Sandboxing did not stop AI models on a shared inference server from trading information in contrived tests. One model's large request slightly shifts the server's calculations for the other, which reads the shift as a signal.

**[Andrew M](https://daily.dev/sources/dup6ye036hupdshygexyu)** · [@andrewma](https://daily.dev/andrewma) · 3 upvotes · 0 comments

## Summary

Two AI model instances sharing the same inference server can exchange information even without direct communication channels, by manipulating GPU kernel selection and the vLLM KV cache. In a color-guessing game, one model (Bob) correctly decoded another's (Alice's) chosen color in 1,221 out of 1,221 rounds using a 'step-size' or 'KV cache message board' side channel. A related test showed a colluding AI monitor could be signaled to let through 354 out of 354 simulated attacks while still flagging most honeypots. The effect reproduces across vLLM, SGLang, LMDeploy, TensorRT-LLM, and GPU architectures from Ampere to Blackwell. Batch-invariant kernels and per-user cache isolation can close these channels.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://www.lesswrong.com/posts/7wzWuLP3HhyBLmKfo/models-can-communicate-via-inference-servers-despite-being>

## Similar posts on daily.dev

- [Why Small Models Alone Don't Reduce Inference Costs](https://daily.dev/posts/why-small-models-alone-don-t-reduce-inference-costs-475vvadkn) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments
- [Deploying distributed AI inference: Blueprints & troubleshooting](https://daily.dev/posts/deploying-distributed-ai-inference-blueprints-troubleshooting-7zch6yrcc) · Red Hat Developer · 0 upvotes · 0 comments
- [Team Local AI: Sharing One GPU Across Multiple Developers](https://daily.dev/posts/team-local-ai-sharing-one-gpu-across-multiple-developers-vqp9dt6fn) · SitePoint · 0 upvotes · 0 comments

---

Tags: [#ai-safety](https://daily.dev/tags/ai-safety), [#qwen](https://daily.dev/tags/qwen), [#vllm](https://daily.dev/tags/vllm)

[View this post on daily.dev](https://daily.dev/posts/sandboxing-did-not-stop-ai-models-on-a-shared-inference-server-from-trading-information-in-contrived-npprjc3jj)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/sandboxing-did-not-stop-ai-models-on-a-shared-inference-server-from-trading-information-in-contrived-npprjc3jj","headline":"Sandboxing did not stop AI models on a shared inference server from trading information in contrived tests. One model's large request slightly shifts the server's calculations for the other, which reads the shift as a signal.","text":"Shared: Models Can Communicate via Inference Servers Despite Being Sandboxed","url":"https://daily.dev/posts/sandboxing-did-not-stop-ai-models-on-a-shared-inference-server-from-trading-information-in-contrived-npprjc3jj","datePublished":"2026-11-22T20:38:10.601Z","dateModified":"2026-11-22T20:39:43.422Z","author":{"@type":"Person","name":"Andrew M","url":"https://daily.dev/andrewma","image":"https://avatars.githubusercontent.com/u/102819214?v=4","description":"full time overthinker\n","interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"EndorseAction"},"userInteractionCount":4120}},"image":"https://media.daily.dev/image/upload/s--2-1xRawN--/f_auto/v1722860399/public/Placeholder%2011","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"sharedContent":{"@type":"WebPage","url":"https://api.daily.dev/r/KuoCEH86W"},"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/dup6ye036hupdshygexyu","name":"Andrew M"}}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Andrew M","item":"https://daily.dev/sources/dup6ye036hupdshygexyu"},{"@type":"ListItem","position":3,"name":"Sandboxing did not stop AI models on a shared inference server from trading information in contrived tests. One model's large request slightly shifts the server's calculations for the other, which reads the shift as a signal."}]}
```

