Vocabulary Attack to Hijack Large Language Model Applications

The fast advancements in Large Language Models (LLMs) are driving an increasing number of applications. Together with the growing number of users, we also see an increasing number of attackers who try to outsmart these systems. They want the model to reveal confidential information, specific false information, or offensive behavior. To this end, they manipulate their instructions for the LLM by inserting separators or rephrasing them systematically until they reach their goal. Our approach is different. It inserts words from the model vocabulary. We find these words using an optimization procedure and embeddings from another LLM (attacker LLM). We prove our approach by goal hijacking two popular open-source LLMs from the Llama2 and the Flan-T5 families, respectively. We present two main findings. First, our approach creates inconspicuous instructions and therefore it is hard to detect. For many attack cases, we find that even a single word insertion is sufficient. Second, we demonstrate that we can conduct our attack using a different model than the target model to conduct our attack with.

Reddiment: Eine SvelteKit- und ElasticSearch-basierte Reddit Sentiment-Analyse

Reddiment is a web-based dashboard that links sentiment analysis of subreddit texts with share prices. The system consists of a backend, frontend and various services. The backend, in Node.js, manages the data and communicates with crawlers that collect Reddit comments and stock market data. Sentiment is analyzed with the help of Vader and TextBlob. The frontend, based on SvelteKit, provides users with a dashboard for visualization. The distribution is carried out via Docker containers and Docker Compose. The project offers expansion options, e.g. the integration of cryptocurrency rates. Reddiment enables the analysis of sentiment and share prices from subreddit data.