Drei Fragen und Antworten: Anger in the KI-System

Wenn productive artificial intelligence, Einzug stopped in Geschäftsanwendungen, kommen auf Unternehmen ganz new Gefahren zu. Researched Sprach Models with Social Engineering and a study on Rollback Augmented Generation (RAG) was conducted and Prof. Dr. Kurz was interviewed by Patrick Levi. Der neuen iX erscheint ein Beitrag von Professor Levi zu Angriffen auf RAG-Systemen.

Anzeige

KI-Chatbots or Large Language Models are generally available in one version, in Volksport they are fast and often fun. Do you know that you can perform jailbreaks quickly, achieve this and more?

Jailbreaks are very enjoyable, with new information for a better experience – and a text for a text card. At this moment, new Anwendungen for large Sprachmodels, can’t sie zur Gefahr werden. You can then use a Chatbot to learn more or learn more in a Chatbot.

KI-Systeme’s Angrife cleverly combines with Social Engineering. Man versucht, sein Gegenüber auszutricksen. Herkömmliche Angriffe portrays the best Schwachstellen. Social Engineering is Angreifer vorteilhaftes, für den Angegriffenen in a way that can have fatal consequences.

Mittlerweile, Rollback Augmented Generation (RAG) is believed to be a technology and it is not possible to use the models in any training and obtain new data. What happened and was this a special Gefahren?

I am Grunde, Chatbot, an Internet provider in the RAG-System; With the support of the offices, I am equipped with Notifications, Emails and other data.

RAG-System’s interest is a system that is often harmonized with texts and is also an Art of Data Poisoning. The Nature of the RAG System also includes: In the RAG System at Zweck, Emails can be sent and different texts can be created between emails. Andernfalls was part of the System. The fury of art that can be best sustained is not realized in normal Jailbreaks and never.

So what happened, um sich davor zu schützen?

At first, Zweck zuschneidet in the RAG-System or the entire KI-Anwendung genau, den sie erfüllen soll und diesen Einsatzzweck möglichst eng. And from now on, further work needs to be done, interacting with Nutzern, API-Zugänge, Datenbank-Schnittstellen and even more.

A large number of features have been made smart with productive KI functions, making it a good method and achieving much better results. This is the analogue of Red-Teaming-Ansätzen in the IT Security class and may be identical to Schwachstellen’s KI-System.

Dear Prof. Thanks to Levi, Dank für Ihre Antworten.

In the new iX and Titel-Artikeln, with the larger Sprachmodels Angriffe, with the RAG-Systemen exclusive Gefahren and their return to the geforderten Sicherheitsmaßnahmen of the AI ​​Act, vorbereitet. Die iX 1/2025 Heise-Store and I’m at the Kiosk.

(ulw)