Multimodal Prompt Injection via Image and Audio Inputs
Attackers can hide commands in images and audio that AI models execute without catching them.
Solène Marchetti
Contributing Writer
Solène Marchetti
A former academic researcher in natural language processing, Solène left a postdoctoral position at a European university in 2022 to report on how language model internals translate into real-world attack surfaces. She brings a technical depth to vulnerability analysis that is rare in security journalism.
4 stories
Evaluating Enterprise Platforms for Indirect Prompt Injection Detection
Most enterprises deploy AI agents without defenses regulators will soon require.
Solène Marchetti
System Prompt Extraction via Jailbreak Sequences
Researchers show how attackers extract hidden AI instructions through conversation.
Solène Marchetti
Sandbox Escape Techniques in LLM Code Execution Environments
Research reveals how LLM sandboxes fail when code execution becomes probabilistic and unreviewed.
Solène Marchetti