SpAIware & More: Advanced Prompt Injection Exploits in LLM Applications

An analysis of advanced prompt injection techniques targeting major LLM applications, including data exfiltration, tool invocation, and ASCII smuggling.

This talk explores advanced prompt injection exploits targeting widely used LLM applications, including Microsoft Copilot, Google Gemini, Google NotebookLM, Apple Intelligence, GitHub Copilot Chat, Anthropic Claude and others. Using real-world demonstrations, we will discuss the following threats in detail: - Misinformation, Phishing, and Scams: Including advanced techniques such as conditional instructions. - Automatic Tool Invocation: Exploiting tool integration to escalate privileges, extract sensitive data, or modify system configurations - Data Exfiltration: Leveraging strategies, such as markdown and hidden payloads, to bypass security controls and leak data. - SpAIware and Persistence: Manipulating LLM memory for long-term control and persistence. - ASCII Smuggling: How LLMs can hide secrets and craft hidden text invisible to users. For each threat category, we will discuss mitigations and show how vendors are addressing these vulnerabilities. By: Johann Rehberger | Security Researcher | Red Team Director, embracethered.com Full Abstract and Presentation Materials: Johann Rehberger | Security Researcher | Red Team Director, embracethered.com