A small number of samples can poison LLMs
From the member
GBTI NetworkAnthropic calls this AI poisoning: planting malicious content in training data so a model learns an unwanted behavior. In one study, just 250 poisoned documents were enough to create a simple backdoor across models of very different sizes.
The contrast with AIO, AI Optimization, is intent. Poisoning tries to manipulate a model secretly, while AIO tries to make legitimate content easier for AI systems to understand, trust, and surface. Both rely on the same basic fact: what gets published online can shape AI behavior.
What are you doing to seed your work into generative AI models?
0 Comments
No comments yet. Be the first. Members comment from the GBTI local client, where comments are submitted as pull requests and auto-published for paid members.
Become a memberComments are for members. Become a member to join the conversation, or if you already are.
You are signed in as a member. .