Indirect Prompt Injections in the Wild – Real World exploits and mitigations Johann Rehberger

With the rapid growth and widespread use of AI and Large Language Models (LLMs), users are facing an increased security risk of scams, data exfiltration, loss of personally identifiable information (PII), and even the threat of remote code execution.

This talk aims to shed light on emerging attack techniques like Indirect Prompt Injections (a vulnerability at the very core of LLM Agents), Cross-Plugin Request Forgery, Data Exfiltration, and more.

The session kicks off with a basic introduction to LLMs, leading to an in-depth exploration of real-world security exploits. We’ll illustrate these challenges using concrete examples and exploits from well-known platforms such as ChatGPT, Google Bard, Bing Chat and Anthropic Claude. The examples will dive into how the attack payloads behind such attacks look like in detail.

The talk will also cover mitigation strategies, and for instance how Microsoft and Anthropic fixed data exfiltration angles reported by the speaker in their Chatbots, providing attendees with practical insights to tackle these cybersecurity issues.

Speaker:

Ekoparty 2023 - HACK THE PLANET
--

Seguinos en la redes: