Project summary

I have a simple hypothesis: As AI agents get more autonomous, we'll need infrastructure that watches what they are doing, instead of only evaluating the models themselves. I want to build the smallest version of that idea and find out if it actually works.

I've aldready done adversial attacks on my system. In one batch I run 233 adaptive attacks, with a strong model as attacker, against my model and Claude. My system catched all attacks and let normal calls pass.

In another batch I run hundreds of iterative tries to cause harmful output with similar results.

I've thus tested if the system discovers attacks.

Now I want to test if my system works to supervide an agent doing tasks, follow what an independent agent does and react if something goes wrong.

I want to build an MVP, as a minimal version of the system, to see if my system works an a supervising infrastructure.

lisaintel.com

x.com/New_AI_Safety

linkedin.com/in/pedro-bentancour-garin-22a0a215

What are this project's goals? How will you achieve them?

Build a simple MVP and test my system.

The long-term goal of my complete projext is to build a global AI governance and safety layer. I belive I have come a ling way with the architechture and the patents I've filed for it. I'm now in a testing and fund raising phase.

How will this funding be used?

Compute and three months runway.

Who is on your team? What's your track record on similar projects?

I'm a PHD candidate at Stockholm university and a serial founder, founded Treehoo. com in 2008, read about that project (from some fans of mine) here:

https://blog.aspiresys.pl/technology/will-treehoo-manage-to-compete-with-google-meanwhile-saving-our-planet/

What are the most likely causes and outcomes if this project fails?

I would need to fine-tune and adapt my system.

How much money have you raised in the last 12 months, and from where?

0\