New research with the UK @AISecurityInst and the @turinginst：

We found that just a few malicious documents can produce vulnerabilities in an LLM-regardless of the size of the model or its training data.

Data-poisoning attacks might be more practical than previously believed.