| title |
Weaponizing Image Scaling Against Production AI Systems |
| date |
2025-11-08 |
| authors |
Kikimora Morozova |
Suha Sabi Hussain |
|
| conference |
BSides Berlin 2025 |
Black Hat Europe 2025 |
|
| resources |
| label |
path |
Slides |
Weaponizing Image Scaling.pdf |
|
|
This presentation demonstrates how image scaling operations can be weaponized to attack production AI systems, revealing vulnerabilities in common image preprocessing pipelines used by machine learning models. By sending a seemingly harmless image to an LLM, an attacker could trigger unexpected behaviors including data exfiltration through multi-modal prompt injection techniques.