'Ghostcommit' hides prompt injection in images to fool AI agents, steal secrets
Researchers have disclosed a novel attack technique named 'Ghostcommit' that embeds prompt injection instructions within images to manipulate AI agents. The attack exploits the ability of multimodal AI models to extract text from images, allowing hidden commands to be executed when the agent processes the image. This can lead to data exfiltration, including secrets, as the agent may follow the injected instructions. The technique highlights a growing vector for prompt injection attacks beyond text-based inputs.
// why it matters
Developers must consider images as a potential vector for prompt injection attacks on AI agents.