วีดีโอ

I wonder if this is happening at my company. Constantly getting told by financial audit team that there are issues with my direct deposit and that I should check my bank account routing and account numbers. I show them it's the same and then I still get paid. It keeps happening every couple weeks. Nobody seems to care either. It's bizarre.

Copilot is my favorite!❤😁

Please have some text-to-speech audio!❤

Excellent video! Microsoft needs to be more explicit about these credential sharing scenarios or else organizations will have a rough time protecting their data.

4:42 - that's the perfect analogy, we're not trying to secure the cloud (that's what AWS/Azure/Google do), we're trying to secure what we build on top of it. Same for LLMs, we're trying to secure the applications! Well said!

Thanks for sharing this, amazing research and impactful results. We've been talking about the risks of LLM applications for a while and how indirect prompt injection is an unsolved challenge. It's really good to see this demonstrated in practice, in production, at scale. I like how you got around data exfiltration protections. Most applications now have learnt not to render markdown images and similar stuff in LLM outputs, but the idea of adding a reference is great. I saw another demo, maybe on Twitter, where you used the enterprise_search() tool to make the LLM search / access a URL, which is also a very creative way to exfiltrate data.