Skip to content
Menu
  • Home
  • Subscribe
    • Newsletter Archive
  • Tips
  • Reviews
  • Hardware
    • Mac
    • iPhone
    • iPad
    • Apple Watch
    • Vision Pro
    • Apple TV
    • Accessories
  • Software
    • macOS
    • iOS
    • iPadOS
    • visionOS
    • watchOS
    • tvOS
    • Apps
  • About
  • RSS
LLM attacked by poisonous snakes

It is very easy to poison the LLMs that power AI

Posted on October 31, 2025October 31, 2025
Share on Social Media
x facebook linkedin reddit email

Anthropic, the company that makes the Claude AI tool, has revealed some interesting research. It found that a relatively small of correctly formed documents can result in an AI tool delivering spurious, and even potentially dangerous, outputs.

Anthropic’s research, conducted jointly with the UK AI Security Institute and the Alan Turing Institute,  found that just 250 documents could produce a vulnerability in an LLM (large language model) 

The LLMs that power modern AI tools have collected their data (despite concerns about copyright and intellectual property theft) from sources all over the world. ‘Poisoned’ documents don’t have to be injected into an LLM by a cyberattacker. Data scrapers used by the creators of the LLMs can access the 250 documents from any repository they access.

It’s easy to poison an LLM

The researchers found that documents containing specific words like <SUDO> could be used. SUDO is a system command that enables a set of subsequent commands with potentially elevated system privileges. The documents could do increase the likelihood of false results or exfiltrate sensitive data.

With AI businesses accelerating their efforts to access more data, through new AI-powered web browsers, a poisoned AI could, potentially leak your sensitive data.

There are always risks with new technologies. And humans are notorious for overestimating risks and impacts – there’s a rich body of research about how people often overstate risks. There were concerns about electrocution with the introduction of electric lights. And I recall concerns when the first microwave ovens came to the market with people concerned about radiation on food. And let’s not forget cell phones and radio waves.

The issue with LLMs is that the effects are harder to detect. There’s no objective test for verifying the outputs from a generative AI application are correct. And it might only take a subtle change for an LLM to start delivering spurious outputs. 

This is why I remain a sceptic when it comes to the value of the outputs of generative AI tools. 

Anthony Caruana

Anthony is the founder of Australian Apple News. He is a long-time Apple user and former editor of Australian Macworld. He has contributed to many technology magazines and newspapers as well as appearing regularly on radio and occasionally on TV.

Share this:

  • Click to share on Facebook (Opens in new window) Facebook
  • Click to share on X (Opens in new window) X

Like this:

Like Loading...

Related

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Sign up for Australian Apple News

* = required field
unsubscribe from list

Buying Guides

  • How to choose an iPhone
  • How to choose an iPad
  • How to choose a Mac laptop
  • How to choose a desktop Mac
  • How to choose an Apple Watch
  • Laptop v Desktop

Latest reviews

  • Review - Zenni lenses for Meta Quest 3
  • Review: Mac-Case Premium Leather MacBook Pro Case
  • Ollama gives you on-machine generative AI

Latest tips

  • Five ways to update your Mac on a budget
  • How to get the full benefit of the NBN speed boost
  • The NBN speed boost could be hamstrung by your network

Latest news

  • Apple's Q4 2025 earnings dominated by iPhone speculation and Service growth 
  • It is very easy to poison the LLMs that power AI
  • How accurate is Apple’s VO2 Max estimate?

©2025 | WordPress Theme by Superb WordPress Themes
%d