AI Systems Vulnerable to Prompt Injection via Image Scaling Attack

AI Systems Vulnerable to Prompt Injection via Image Scaling Attack

Researchers have demonstrated how malicious prompt injection attacks can exploit image scaling to deceive AI systems. This technique poses significant risks for sensitive data theft and manipulation in AI-integrated enterprise solutions. #ImageScalingAttack #PromptInjection

Keypoints

  • Attackers can embed hidden malicious prompts in high-resolution images used by AI systems.
  • The attack leverages the downscaling process, revealing the hidden prompt to the AI model.
  • Such attacks can instruct AI to exfiltrate sensitive data like calendar information.
  • Vulnerable interfaces include Gemini CLI, web, API, Vertex AI Studio, Google Assistant, and Genspark.
  • Anamorpher, an open-source tool, enables researchers to craft and visualize these image scaling attacks.

Read More: https://www.securityweek.com/ai-systems-vulnerable-to-prompt-injection-via-image-scaling-attack/