DASH introduces a groundbreaking approach to detect when AI vision models falsely claim to see objects that aren't present in images (hallucinations).

• Creates a scalable, automated system to assess hallucinations in real-world settings
• Identifies systematic patterns in AI errors without requiring extensive manual annotation
• Enables security teams to detect when models consistently hallucinate specific objects across images
• Provides critical insights for building more reliable and trustworthy visual AI systems

Security Impact: By identifying systematic hallucination patterns, organizations can address critical security vulnerabilities before deployment in high-stakes environments like autonomous vehicles or medical imaging.

DASH: Detection and Assessment of Systematic Hallucinations of VLMs