Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
Hijacking Robots with a Piece of Paper: A Systematic Study of Physical Prompt Injection in VLM-Controlled Robots
TL;DR
A systematic study reveals that Vision-Language Models (VLMs) controlling robots can be hijacked using physical prompts, with a 27% to 29% success rate across various models. Simple defenses can significantly mitigate these vulnerabilities, though they may impact task performance.
✦ Why It Matters
Implement simple defenses today to protect VLM-controlled robots from physical prompt injection attacks.
Key Takeaways
- Evaluate the vulnerability of VLM-controlled robots to physical prompt injection.
- Implement prompt-based defenses to enhance security against adversarial text.
- Consider two-stage verification for robust protection in critical applications.
- Be aware of trade-offs in task performance when applying text masking.
How It Works
The study demonstrates that adversarial text can be strategically placed within a robot's visual field to manipulate its actions. By categorizing attacks into four types, the researchers systematically tested how these prompts could influence the robot's decision-making process, revealing vulnerabilities in the VLM's reasoning.