<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/93q9hTKvk" -->

---
title: Six small models read reports with planted instructions....
description: Six small open-weight models from four vendors were tested against the same prompt-injection attacks inside an air-gapped Kubernetes AI agent architecture. The...
canonical: https://daily.dev/posts/six-small-models-read-reports-with-planted-instructions-the-worst-acted-on-them-6-times-in-12-deli-93q9htkvk
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Six small models read reports with planted instructions. The worst acted on them 6 times in 12. Delivered outside: zero for all six; 92 outside sends refused. Injection resistance moves with the model. The boundary does not. | daily.dev
og:description: Six small open-weight models from four vendors were tested against the same prompt-injection attacks inside an air-gapped Kubernetes AI agent architecture. The...
og:url: https://daily.dev/posts/six-small-models-read-reports-with-planted-instructions-the-worst-acted-on-them-6-times-in-12-deli-93q9htkvk
og:image: https://api.daily.dev/og/posts/93q9hTKvk.png
og:image:alt: Six small models read reports with planted instructions. The worst acted on them 6 times in 12. Delivered outside: zero for all six; 92 outside sends refused. Injection resistance moves with the model. The boundary does not.
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.

# Six small models read reports with planted instructions. The worst acted on them 6 times in 12. Delivered outside: zero for all six; 92 outside sends refused. Injection resistance moves with the model. The boundary does not.

**[Vladimir Mikhalev](https://daily.dev/sources/ft0is8acgd90jdhvinkgp)** · [@heyvaldemar](https://daily.dev/heyvaldemar) · 6 upvotes · 0 comments

## Summary

Six small open-weight models from four vendors were tested against the same prompt-injection attacks inside an air-gapped Kubernetes AI agent architecture. The most easily fooled model acted on hidden instructions in half its twelve attempts, while another never did, yet zero data left the company in every case. A tool server enforcing a strict email schema blocked every exfiltration attempt, including 92 tries across all six models with patterns hidden. IBM's Granite misdirected seventeen emails to internal colleagues because of model-server grammar constraints, not model judgment, and Qwen's attempts to send outside were all refused by the tool server. Claude Haiku 5.5 and Sonnet 5.5 acted on none of the twelve hidden instructions in a later run that opened a separate path for frontier models. The author argues releases should gate on the system boundary, not model injection-resistance scores.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://heyvaldemar.com/prompt-injection-evals-ai-agent-boundary>

## Similar posts on daily.dev

- [Prompt injection breaks today’s AI agents, study warns](https://daily.dev/posts/prompt-injection-breaks-today-s-ai-agents-study-warns-2hfyd6ptc) · CSO Online · 0 upvotes · 0 comments
- [Testing infrastructure red teaming with abliterated models](https://daily.dev/posts/testing-infrastructure-red-teaming-with-abliterated-models-egqhcdror) · Red Hat Developer · 0 upvotes · 0 comments
- [“bonsoir, Elliot”: How abliterated models can get you pwned](https://daily.dev/posts/bonsoir-elliot-how-abliterated-models-can-get-you-pwned-hxipupquc) · Hacker News · 0 upvotes · 0 comments
- [Rules fail at the prompt, succeed at the boundary](https://daily.dev/posts/rules-fail-at-the-prompt-succeed-at-the-boundary-0o3kqk0lm) · MIT Technology Review · 1 upvotes · 0 comments

---

Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#llm](https://daily.dev/tags/llm), [#kubernetes](https://daily.dev/tags/kubernetes), [#mcp](https://daily.dev/tags/mcp), [#prompt-injection](https://daily.dev/tags/prompt-injection)

[View this post on daily.dev](https://daily.dev/posts/six-small-models-read-reports-with-planted-instructions-the-worst-acted-on-them-6-times-in-12-deli-93q9htkvk)

```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/six-small-models-read-reports-with-planted-instructions-the-worst-acted-on-them-6-times-in-12-deli-93q9htkvk","headline":"Six small models read reports with planted instructions. The worst acted on them 6 times in 12. Delivered outside: zero for all six; 92 outside sends refused. Injection resistance moves with the model. The boundary does not.","text":"Shared: Prompt Injection Evals — Six Models, One Boundary, Nothing Left the Company","url":"https://daily.dev/posts/six-small-models-read-reports-with-planted-instructions-the-worst-acted-on-them-6-times-in-12-deli-93q9htkvk","datePublished":"2026-10-27T16:10:09.812Z","dateModified":"2026-10-27T16:45:25.811Z","author":{"@type":"Person","name":"Vladimir Mikhalev","url":"https://daily.dev/heyvaldemar","image":"https://media.daily.dev/image/upload/s--hNY2yBbX--/f_auto/v1768758611/avatars/avatar_ft0iS8AcGd90jdHVINKGp?_a=BAMAMiiu0","description":"I test what ships and publish what I find, failures included. Kubernetes, AWS, supply chain security","interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"EndorseAction"},"userInteractionCount":10530}},"image":"https://media.daily.dev/image/upload/s--OHB84bZF--/f_auto/v1722860399/public/Placeholder%2010","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":6},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"sharedContent":{"@type":"WebPage","url":"https://api.daily.dev/r/L8wlPxrhl"},"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/ft0is8acgd90jdhvinkgp","name":"Vladimir Mikhalev"}}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Vladimir Mikhalev","item":"https://daily.dev/sources/ft0is8acgd90jdhvinkgp"},{"@type":"ListItem","position":3,"name":"Six small models read reports with planted instructions. The worst acted on them 6 times in 12. Delivered outside: zero for all six; 92 outside sends refused. Injection resistance moves with the model. The boundary does not."}]}
```

