[Webinar] Building LLM applications in a secure way (WithSecure™)

The webinar begins with Yanen introducing expert Donato Capitella to discuss large language models (LLMs). They highlight the surge of interest in LLMs like ChatGPT and open with a conversation about their origins and the advancements in machine learning that preceded them.

Donato elaborates on the history of language models, explaining that the early challenges involved translating sentences from one language to another and how this led to innovations in the encoder-decoder architecture over the past decade. He then introduces the attention mechanism which has become crucial in modern NLP tasks.

The discussion transitions into the introduction of Transformers, developed by Google in 2017, which combined previous advancements in encoder-decoder models and attention mechanisms, thereby allowing for effective parallel processing of data important for training sophisticated LLMs.

The application of LLMs is discussed, defining applications as wrappers around the language models for specific use cases. Examples given include the utilization of LLMs for generic tasks like summarization or the more specific use cases like enhancing features in services like Google Docs.

Donato shares his viewpoints on promising future applications of LLMs, targeting the development of autonomous agents capable of interacting with their environment and performing tasks autonomously without human oversight, hinting at the potential evolution of embodied AI.

The safety of LLM technologies is addressed, specifically referencing misconceptions surrounding their security risks likened to the science fiction notion of 'Skynet'. Donato insists that current LLM technologies lack true consciousness and discusses potential security challenges they face.

Yanen prompts Donato to showcase a demonstration related to LLMs and risks. The demo illustrates a recruitment scenario where an attacker exploits a job application through prompt injection to extract confidential candidate information from the system.

Demonstrating how LLM applications can be exploited, Donato illustrates how attackers can leverage vulnerabilities within LLM systems—most notably through prompt injection, which manipulates LLM outputs to exfiltrate sensitive information.

The discussion expands into potential real-world threats presented by LLMs, showcasing various examples of prompt injection risks, followed by emphasizing the importance of treating LLM output as untrusted and the need for strict security measures to handle these vulnerabilities.