Date: June 11, 2024 

Time: 3:00 pm to approx. 4:30 pm 

Location: DFKI Saarbrücken, Room Reuse HG 2.17 

The event is organized by the Center for European Research in Trusted AI (CERTAIN) and offers the perfect opportunity to learn more about this aspect of Generative AI. There will also be a demonstration as well as room for discussion and questions. 

Compromising Large Language Models (LLM) at large scale  

How to easily hack generative AI 

Large language models (LLMs) are currently used widely and intensively, but are vulnerable to attacks. So far, this possibility has been considered very little, or purely from the perspective of data protection. However, this is not nearly enough; the really relevant threats are elsewhere and it is only a matter of time before they become real.  

Indirect Prompt Injection enables a remote takeover of LLM applications on a large scale. An attacker smuggles hidden instructions into the dialog context of a language model via external sources (websites, documents, etc.) and takes control of the dialog. The user is unaware of this. 

This vulnerability was published and demonstrated by sequire technology in February 2023. Extensive discussions were held with affected providers such as Microsoft, OpenAI and Google. In the ranking of the most dangerous vulnerabilities of language models (OWASP Top 10), prompt injection was listed as the top 1 threat; the German Federal Office for Information Security published a warning based on sequire’s work.  

Speaker:  

Dr. Christoph Endres, born in 1971, is a computer scientist and originally an AI researcher. After completing his doctorate in the field of intelligent driver assistance systems with Prof. Wolfgang Wahlster (DFKI), he switched to the field of cybersecurity and has been Managing Director of sequire technology GmbH, which he co-founded, since 2021. He is part of the team that found and analyzed the Indirect Prompt Injection vulnerability in language models (https://doi.org/10.1145/3605764.3623985).