Riptide - Taking Back the Blockchain from AI Blackhats

Transcript

So, I want to touch on two things. Uh, number  one is I almost got scammed yesterday. So, we released this big news on our end. You could  check my Twitter, Zeroxiptide, and we found we protected 28 million from immediate theft. uh we  were rewarded a quarter million in a bug bounty and so we had a lot of journalists, a lot of  people reaching out and so this person reaches out and she says, "Oh yeah, I'm a Wall Street  Journal journalist." I said, "Okay, wow, great. Okay, let's let's set up a call." And uh she  says, "Oh, uh put in my Calendarly." Okay, great. And then she says, "Oh, here you can you  could download uh our our app here, Sy." And I not downloading anything. What is this? Can we  just do Google Meet? No, no, no. It's uh it's it's

banned. we can't do in our organization. Okay, I  deep dive like this is how good they are is on the journal's website. They had an outdated link from  an old reporter who had left the company, but that ghost link when you when you go recursively  through the directory, it's page not found, page not found. But this this link was still there  and it was legit. And then her supervisor, who was another outdated reporter, her ex account had  been compromised. She said, "Oh, you can verify. Contact her." I contact her. Oh, yeah. Yeah. have  legit. And so I finally put it together. I'm like, "Oh my god, this was like the second time  in a week where I almost got scammed." So it's so real and um it just makes me think of  like another analogy before I kick this off is is like goes into fitness, right? Um I used  to be big into the gym, right? Go to the gym,

think I was great. And then uh during  COVID all the gyms closed, so I said, "Hey, I'm gonna go to the calisthenics bars." And  I went out there and I thought, "Oh man, I know everything about fitness and these guys are  doing handstands on bars. They're doing balancing, all these things that I couldn't do."  And I thought, "Oh man, this is uh, you know, it was an eye openener." Like there's  always something that you don't know about whatever topic you think you're an expert in. So  like security, we're on the blockchain a lot. We think, hey, we know everything about it and then  you could get fished and lose your private key. uh there's you know a web two compromise in the  front end you don't see and you sign the wrong thing. So there's always something that we could  learn and I think AI is one of those things where

uh for bug hunters like me it's it was kind  of hard to accept that it was better than me. Like it could find bugs quicker than me. Yeah,  there's false positives. Yeah, there's there's downsides to it. But the way it could uh ramp  you up to a codebase and explain everything and identify things that you would miss, I  mean is amazing. You have to recognize that and just adapt to the future. That's it.  So, um how do we fight back about this AI black hat threat which is happening? Um there's  like there's no specific statistics, right? But what's happening over the last 12 months and I'll  peg it at last summer because that's when we had

kind of a a very good working protocol of Greg AAI  and we're able to find public bug bounty criticals with it. And so the last 12 months I started  looking at these hacks and I'm seeing things that uh first I noticed an uptick in the frequency  of these hacks which is surprising. You're like okay what's what's happening here? And then  um the complexity uh wasn't really there. You'd see hacks that kind of are basic hacks,  you know, basic bugs that keep resurfacing, but maybe just no one could find them. And so  these AI hacks, people are obviously scanning and they're finding these things and black  hating them and exploiting them. So 1.6 6 bill.

A lot of this is um not just like it's like the  the drift hack that was just explained. Like a lot of this is private key compromise, fishing.  Um there's so many other vectors than just if you look historically at all the hacks, it's always  private key compromise. I mean, the humans are the the weakest link as always. Uh but there's still  a lot of D5 smart contract exploits but more and more we're seeing more on um DT distributed  ledger ledger technology so blockchain back end uh who am I if you don't know me um  found uh probably the biggest bug was uh in arbitum where I took over the  bridge showed that you could take it over uh that was a monster but I've been doing this for  about five years uh hunting bugs manually and now

uh since last summer I saw the shift with LLMs and  And then um me and my co-founders started Grego AI and we basically like plugged in my  bounty hunting brain into um an AI. So all right so what does this look like?  Like you see these black hat attacks and you're like okay what are they doing? Are they  just saying you know find all bugs? Okay. No. Imagine that you could take on chain. It's very  easy if you know what you're doing to steal money and you can take whatever you want. You could  wash it. You could do all this anonymously.

If you screw up, that's on your own obsc.  But you could provably do this unfortunately uh because of the blockchain and and how privacy  works. So these guys, I mean, they've they've automated the whole flow um almost and it's just  about fine-tuning their models um paying the cost for compute. If you have state level actors like  North Korea, that's not really a concern at all. But yeah, I mean, you scope out the chain, you  find targets. A lot of these targets are um contracts that were never verified that no one  bothered to take a look at because they don't have the time to decompile uh to hunt down these  bugs specifically and and you know just take the time to do it with AI it's trivial why not do it  uh crafting the payloads for for expportation um

executing everything like everything  is just so seamless and so easy to do it's I mean this is this is the the kind of  the arena that we play in. So the black hats are going to black hat and so what do we do, you  know, to to defend against this seems impossible. Uh here's what we're doing now. So right  now we have manual code code audits. These point in time audits I think are the biggest  problem that we need to address. We have AI being able to do this so much quicker. Uh there there  are major flaws to it, but what we could do now like why do this kind of um manual audit two three  months in the pipeline? It costs a lot of money when we could do this uh a hybrid audit. I I still  love manual audits. They're great, but we can

we can um add to this using AI using proper AI  techniques and um constantly audit the protocol. There's no reason you can't constantly review  your protocol once a month, once on every PR, or you know, after every major code revision  or update, you should review it. It's not just this is fine. Uh four years later, it's going to  be fine. You just you can't assume that anymore. Uh we still have static analyzers, those are  still good. Um bug bounties, onchain monitoring. So a lot of this is just like every angle. It's  like having like like you have a bank, right?

And normally a physical bank can get breached  by only some really gangster attackers that are willing to risk it all. And then you had the  blockchain where you could simulate your attacks. And now you have a guy amped up on  stimulants who is plugged into the biggest supercomputer in the world and  he's constantly attacking your protocol. And it's just it's insane. And so all the old  techniques you just you have to fight AI with AI. So some of the things I can recommend um  proactively AIT like I was talking about adversarial red teaming. This is this is like  what we used to do what any security focused team uh would do along with the amplified  bug bounty programs. It's like

instead of thinking like a dev, which a lot of  devs do, is like how do I make this flow great for the user? What do I do for the user? You need to  think about what if you were the hacker, like why would you deposit first? Why wouldn't I  withdraw first? Why wouldn't I call the uh uh accumulate interest function? Like just think  of weird things like that. And you should have a a cohort of SRS with your protocol in a war room and  you say, "Hey, test my protocol. Like, does this make sense?" and they're going to try to break it  and they can use AI to break it. They could use whatever they can to break it, but you can't think  like a dev um and just constantly assume that they the current flow and the plan flow everything  is going to go to plan. It never happens.

Okay, these are these are Riptide's  favorite things to tell devs. Uh number one, security first.  you if you are a vibe coder, you must focus on security. If you have no idea  what you're doing in your vibe coding protocol and you're going to put that out there  in the blockchain, you're insane. If you're a dev that knows what they're doing and  you're vibe coding, cool, because you're going to check everything afterwards. But you have to  develop this mindset like how am I going to build this up where it's going to be secure,  not just from a functionality perspective.

uh immutability which I love to see which we see  in uh from from Michael from curve we see with unis swap we see with a few protocols but  everyone is hooked on the proxy contract and they love being able to upgrade they love admin  keys and centralization and everything's we flag as auditors and we look at bug hunters and and we  try to tell protocols about these things and um they just they want that centralized control. One,  for flexibility. If something goes wrong, they want to be able to pause it. They want to be able  to make changes. Uh but two, it's just it's just it's so unsafe and no one realizes this or they're  willing to accept the risk of it being unsafe.

Uh your test suite. This is a  great way to find bugs that um I think a lot of experienced bug hunters use  is they can they can look at your codebase and they could say well what questions did they  ask? Let me go look at their test suite. And this is why some projects won't reveal their test  suite. But you could see exactly what the devs are thinking. And you could say, well, what did they  test for? What what questions are they asking? And if you don't ask the right questions, well,  you're missing things. And when you're missing things, I'm going to try to find those things  and find bugs and try to break your protocol.

Like I said, the black hat mindset. Uh, always  think like a black hat on your protocol. Uh, and this also goes with when you use  AI. So, if you're going to make a a prompt, and there's so many prompts and skills out there,  I can get into all that, but if you're going to make a prompt, make sure you're asking the right  questions on this prompt. Are you going to say you download a pash of skills or or whatever?  Or are you just going to ask certain questions that you know about your protocol and you're  going to say well can this you know if I was a hacker and I come at it very exact. So you  want to describe an exact pathway. You can't just be general and say find all the bugs. You  have to specifically tell it this is the attack I'm thinking of. Would this work? And if you do  that consistently and you write the right tests,

you're going to be a very secure protocol.  I'll tell you that uh you get formally verified by Certora or you use their tool and you think,  "Oh, wow. This is great." That's that's just like your test suite. If your spec is wrong and you  verify an incorrect spec, well, big deal. You've you've done nothing but you've just showed uh that  investors will think you're secure because they see uh that you've been proved, but your spec is  wrong. And there's another bug waiting to pop up. And last one, uh, last two,  Ask AI, like I said, uh, with Greg AI, we have a great product, I think.  Um, we're the only deep scan kind of AI program out there. Uh, but you can also just just use  it on your own. Just ask those right questions.

Don't don't fall into the skills trap and all this  stuff. Just just ask pointed questions at the LLM. Uh the last one is kind of uh point of contention  with some people but I think DeFi in general is absolutely crazy that it's 2026 and you could  drain entire protocol in one transaction still to this day and we wonder why people are  taking their money out of the ecosystem. if you have any rate limiting circuit breakers  something like that like why and I get it right I'm all about DeFi but why why do you have to  withdraw a hundred million in one transaction you can't do that over two three like that's that's  all not possible and I I understand composability flash loans the argument but this is not the worst  thing to have and I do hope we see more of it

uh some things you could use today the typical  things that we have here uh I'm sure you all heard of these Slyther roll. Uh fuzzing. Uh I' I'd  like to give Recon some love. Uh made by uh Alex. Uh really good framework for fuzzing.  Uh the built-in foundry stuff. Akidna. Uh if you don't use Tenderly, they're  really good for doing your simulations. Uh especially transaction simulations  on anything you're going to sign. Uh build your own skills, your own harnesses using  AI. Um just your protocol is unique. You can even use AI to develop your own skills and harness,  but customize it specifically for your protocol.

Uh, and my favorite, [clears throat] your  brain and your favorite SR, right? These are the final kind of human things because I advocate  a hybrid audit approach, hybrid security approach. Like read your code. I don't care how big  it is. If you're going to attract major TVL, you need to read your code. And this is what  all the top any any good audit firm will do this line by line. and you're going to review your code  and just focus, think about it from an adversarial mindset and then get your favorite uh hacker in  there, incentivize them, give them some future tokens and have them try to break your code all  the time. Uh humble shill, right? So at Grego AI, we built a couple things. First was our deep  audit and this was this was uh the whole goal of this company right was to find bugs that humans  can't find because LLMs are probabilistic they're

very different than static analyzers uh they're  different than anything else because you've seen this you could run a prompt and you could say okay  it gives you five bugs you run the same prompt you get five new bugs and you're like this is this  is really crazy you do this on complex code bases and when you when you use our framework,  it's it's crazy what it comes up with. Um, and it's very interesting in some of  the edge cases that it comes up with. As a bug hunter doing this professionally, I would  have never combined the different bug classes in a way that it has. And that's why I see  it as like you have to have a hybrid audit.

you can't just do human alone because there  are bugs that any top auditor will say yes we we missed it and that's our eye openening  moment to use AI. So we have kind of a a cheaper version which is a surface scan. Uh we would  say to use that after a security review. Uh the deeper audit is like you're about to go  mainet. We compare it to a human audit. It's not this isn't uh like the other AI scan competitors.  It's much more compute heavy, but the whole goal is to promise security for your protocol. Uh, so  if you've seen the news recently, we protected 28 million. Um, we can't disclose the protocol uh  earned a quarter million in bug bounty uh just for that one. We've earned a half a million  uh in bug bounties over the past few months.

Uh we have we've been building this since um  before uh summer of 2025. So, it's been um a pretty pretty deep uh deep build in the making.  And my co-founder is is big into AI, younger guy, and we combined kind of his AI knowledge with  my bug hunting knowledge, and we created this this crazy project to hopefully help enhance the  security of Ethereum and all these EVM chains. It's language agnostic. We found bugs in in Go  and Move in Solidity. uh it doesn't really matter. But the whole goal is like I can't  realistically and and I'm a OG blockchain, you know, Bitcoin, Ethereum guy.  Uh and I I believe in it and I think that

this is the future, but we have so many security  problems that we need to fix. And it's I can't even recommend it to investors because security  is a joke, unfortunately. So, I'm trying to do my part while building a business to help hopefully  improve the security of the ecosystem. [sighs] So, you can price out a scan uh for yourself if you'd  like. Scan.gragoai. It's still kind of whitelisted because I can't have black hats getting on there  and paying for some scans and then we give them the mugs. Um, but feel free to try it out.  Uh, if you want more info, you know, please uh please message me and and we could talk  further. So, that's all I got. Thank you.