Author(s): Alok Choudhary

Originally published on Towards AI.

LLM Poisoning!

LLM poisoning is a way to deliberately corrupt an AI model by feeding it a small amount of bad data during its training. The core goal of a poisoning attack is to make the AI give incorrect, “gibberish,” or pre-determined answers whenever it sees a specific trigger — effectively creating a hidden, malicious instruction that the model feels compelled to follow. A recent discovery by researchers from the UKI Security Institute, Alan Turing Institute, and Anthropic has shown this is much easier to accomplish than previously thought.

The article explains the concept of LLM poisoning, which involves embedding malicious commands in an AI’s training data to manipulate its behavior. By using few corrupted documents, attackers can induce significant errors in AI outputs. The research emphasizes that only 250 malicious entries are enough to compromise even large AI models, challenging previous assumptions about the scale needed for such attacks. It reveals the subtlety of these attacks and highlights the ease with which adversaries can corrupt AI systems.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI