Security Considerations of DeFAI

Transcript

after noon. Um So, my name is Victor Okafou and I'm a security researcher at QuillAudits and I'll be speaking on the security consideration of Def AI. And I want to thank you all for being here. Uh one stuff I want to mainly share on is that the Def AI ecosystem just got a lot bigger and it's mainly because of, you know, Def AI. And basically what it is is the integration of the AI automation with the, you know, decentralized financial system. But this new innovation has taken the world by a storm. Like the AI agents are no longer just replying to it, but making real financial decision and performing complex multiple Def AI operations on behalf of non-technical

users. And so in summary, Def AI is amazing. Right? But there's a problem. So, the thing is that you see this new innovation that, you know, have taken the world by a storm can do amazing stuff. Also introduce new security risks and new threats. And so, we need to be asking how do we secure Def AI? And the truth is that when it comes to securing Def AI, it's different from normal security. They have to consider three key layers.

And this layers is the data, which is what we use to train and tune our model. And then we have to then consider the model, which is the brain itself, which is what we interact to get the output, the usage. So, let's start with the data. So, basically, um so, the model the accuracy of the model is highly important and it's dependent on the data itself. So, one stuff that people could do is to look at the data and see what they could do to it because that's what they do. If something important exist, they try to break it. And one way they break it is by poisoning. So, poisoning is where you try to impute, you know, inaccurate data into the data system and cause it to do the wrong thing. And we have cases of like having oracles been tricked to perform a trade with the wrong data.

So, there's also cyber attack and exfiltration. Exfiltration is similar to poisoning, but it's almost the it's almost the opposite. So, exfiltration you're trying to get data out of the system and not really trying to input data. And so, yeah. How do we protect the the data? So, what we actually kind of want to be doing is classification. So, you want to know So, I want to know where the right data is and the ones that are not right, I want to know where they are. And I also want to also encrypt the data sets so that in case of an exfiltration attack, like I said, exfiltration is where you try to get data out of system, you want to ensure that it won't cause any harm because you know that it only requires authorized users to actually read the data.

And we need to strengthen access control and monitoring. So, now let's talk on the model itself. So, model is uh the brain itself, but it operates kind of differently because most people don't actually build the model from scratch. Uh the reason is because it's very difficult to make and it's expensive. So, what they do is they use existing models and you could be integrating a model that you think is trusted, but then you find out that it's actually the fake one or a copy of the model. You have a case like that happening and then you get hacked. So, you also want to watch out for the API parts as well. Uh that's also one way you could get into the system and cause a hack. And lastly, there issues of plugins having hidden privilege rows and

you want to check out for that and also legal risks and copyright issue. So, how do we secure the model? We want to verify. You want to verify the source. You want to know it's actually the model you're integrating and you also want to scan for any hidden privileges. And that's where you want to have a separate wallet keys for each API uh AI you try to integrate because so that in the times of an attack it doesn't escalate around. And of course, ensure IP be So, now let's talk on the usage this time around. Funny enough, the usage currently right now is the usage that has the highest point of attack. And one of the most common one is prompt injection. And prompt injection usually leads to biased output, and it leads to like denial of service attacks.

Basically, where you overload the AI with so many prompts, it starts slowing down for quick transactions. And there's also model attacks. That's where you kind of craft malicious queries to make the model give you strategy of itself. So, and lastly, you want to define how to secure the usage. You want to monitor for the red patterns and queries. And you also want to implement guardrails and restrictions of certain things the AI should never do. You want to make sure it's implemented. And lastly, you want to have detection. So, yeah, securing the DeFi AI means securing the future of finance because, to be honest, DeFi will always be complex. So, you will not you won't actually reach the stage of mass adoption at the market.

So, one way is actually through the institution, and the second is actually DeFi AI. So, yeah, thank you.