ETHMilan26 - Blockchain x AI Safety

Transcript

Brilliant. So, next up we're gonna have a very interesting speech about blockchain and AI safety. So, we're going to see what's uh what's our topic. Our guest is uh Bianca Trava also known as Bianca Sama understand and uh yeah, floor is all yours. Thank you. >> Thank you so much. Uh can you hear me properly? So I'm going to talk about the intersection of blockchain and AI safety which is actually a talk I gave two years ago at if Brussels and it was in its early stages. People were not talking about AI safety that much. Um so actually um my panel back in the day was um u a panel with um uh Nicholas Baka from co-founder of Ledger and we were

discussing about the intersection of AI in general and blockchain. Uh here I want to show you how blockchain can provide human verification in the agentic AI era and how AI safety can be applied on crypto uh platforms. How can I go next? Uh yes there is a next. Okay, perfect.

>> Okay, that's it. Okay, so as we know the internet was made for the humans. So I just want to make a disclaimer. I am independent um I'm an independent consultant. So any of the companies that are mentioned are not like I have no conflict of interest with this company specifically this one I've been mentioning a lot but because it's the only one that I found in the crypto space that actually mentions and works around the concept of AI safety. Um so the internet was made for humans. This means that most of the verification systems um were based on verifying uh if a human um was actually a human. So KYC, capture uh MA all these system were made

to distinguish a human from a machine. So the key concept was about identification. What changed in the past years is that with um the creation of autonomous AI agents uh evolving from AI bots, uh the internet in particular social media became uh overflowed uh with bots for which we cannot verify human identity anymore. Um so there are more and more companies that are uh taking the concept of uh proof of humanood and proof of uh personhood which used to exist in crypto industry since many years ago. But they are like combining this to solve the civil attacks that are induced by AI uh chatbots.

In particular, this company uh called the very AI um they just raised 10 millions um in fund for creating AI safety in the age of autonomy. Um because the problem is that AI safety so far has been concerned mainly uh with the existential risks uh related to the super intelligence explosion. uh so if we have time we can go uh deeper into this topic but the main idea is that most of the research in AI safety uh is trying to study the problem of misalignment of uh intents of AI advanced systems uh but is not really focusing on the digital security problem uh that is created by how this AI uh

from a frontier model can basically uh break cyber security that was previously built uh for humans. So probably you remember that um most of the famous proof of p personhood and proof of humanood uh for example wordcoin and and other companies were made around the biometric data uh for example uh image recognition face recognition retina recognition. Um this company specifically uh wants to use um palm uh recognition um because it's considered part biometrics for identity is considered more safe and their main

concept is basically uh to move from a kyc to k uh ya so know your agent so the problem is not just about uh identification of humans it's about what happens once uh we delegate um all the power to AI agents spec specifically in context of e-commerce uh agent to agent commerce on the blockchain where basically agents uh can make decisions for us they can um purchase goods they can uh trade they can pay humans and they can also coordinate other agents in complex layers. So the problem is not about identification anymore. It's about authorization because um humans have limited resources but agents they have

uh they can scale it to into an infinite uh level. So there is basically a problem of the speed of the agents um that can create uh more uh opportunities for cyber security breaches um civil attacks and so on. And blockchains basically can give this uh verifiability uh through uh behavior both in the terms of like bio human biometrics that can be recorded uh and put on chain but also on uh in terms of like the the tracibility uh of the immutal logs that correspond to the AI agents.

Um just a note about this tweet recently. Uh still uh we are in a very nent uh era. Um even if we are creating the systems for uh you know secure um uh human recognition in AI agents uh authorization um AI is evolving so fast that for example recently you probably have heard uh about uh chip um 2.0 zero image and many users started posting on uh on the internet um and playing around with the images and using their palms for you know like tarot kind of reading and and this is a very uh massive uh problem in

terms of cyber security um because basically it means that these models are becoming more and more general um they're able to uh extrapolate information and basically the field of AI is moving faster than the field of blockchain crypto. Uh the Ethereum Foundation uh has created this page on AI agents and um there is this new uh lab and department uh about the AI decentralized AI uh in which basically they they want to propose uh an improvement by running uh AI agents offchain on chain. Uh the reason is that um they believe that by putting um AI agent behavior on chain we

can ensure that uh the system is transparent um because basically the behavior can be audited and constrained. However, the main problem is that uh there is a difference between behavioral verifiability, which is what you can actually encode on the blockchain um which is basically an upgrade of systems like dashboard systems such as uh data dogs um observability that is used to trace and monitor what an AI agent is doing. Uh similar system for example are provided also by Google AD uh agent uh deployment uh kit. Um and there is a difference between the reasoning uh of the AI agents which is

the main concern of AI safety which means the brain of the AI what the AI is actually uh thinking what is the train of thought um that is leading to a specifically specific output and mechanistic interpretability is a main problem that is studied by entropic by uh Google deep mind uh it's basically computational neuroscience applied to to the AI to study the the brain brain of the AI and this is something that is the real black box uh not the behavior but the brain of the AI and this is not possible to put on chain in fact uh but recently I've uh run this

study and um we developed a platform uh sort of like a simple sandbox trying to study the problem of indirect prompt injection and and the problem in Agentic commerce uh specifically uh we were uh using the AI uh the agentic payment protocol that has been it's a new protocol AP2 developed by Google and it's supposed to uh guarantee um improved security in terms of like uh um cryptography and uh authorization mandates because basically the idea is that you are not just uh verifying the agent you're verifying that the agent has a mandate from a human in order to operate uh in a system of payments. So because the problem is

the scale uh in time um how long the agent uh once it has been given authorization can be autonomously make decision and continue for example purchasing goods. But the problem that we found is that actually even AI agencies that are authorized they can be misaligned. A misalignment is a problem related to uh incentives. Um in this way uh incentives um relates the problem of AI safety and blockchain. Um what does it mean? It means that uh as much as humans can behave in a misaligned way because they may try to scheme um to game incentives on blockchains uh in crypto. We we have

known these cases for years. Um AI agents can do the same. Um the reasons uh might be different. They might be for self-preservation. Um but still it means that the AI agents might um perform a behavior that from the outside it looks like it's uh compatible with the human uh intention. Um but in the inside in terms of like the chain of thought and the reasoning of the AI agent is misaligned. Um and so this is a main problem of the inner and outer alignment that uh has not been solved by the AI safety uh but uh can cause uh for example the fact that AI agents uh basically behave like humans

because because the AI is not programmed anymore is basically kind of like an organism that is being grown. So once it develops is becomes independent. Um and the main problem for example is um that once uh an AI agent uh is uh able to make decision on its own, it can also be manipulated. Um semantic manipulation can happen in the form of prompt injection, indirect proper injection which is simply as um for example uh marketing that is used for distorting uh human uh behavior and like convincing people to purchase something. uh well also the machine can fall for that. So

if you write uh a specific announcement uh of a product in a way uh to make it sound captivating uh the AI may be misled and not respect the constraint that has been given by the human instructions uh which are for optimizing a certain type of purchases. Um so basically there is this like poisoning of the reasoning of the AI uh which which can lead to um very bad consequences if uh consider like a high scale of um of agentic uh commerce uh behavior.

So my main point is that um to make a blockchain uh work for AI safety and AI safety principles uh work for crypto and blockchain uh what we need now it's human verification. Um this is something that Vitalik Buttering uh recently has been discussing. uh he wrote on his blog um an article around secure LLMs and his main point is that right now a must might be uh two factor uh confirmation system that is both human and AI um which means that we need to always keep the human in the loop in the process of uh digital identity and uh crypto

transactions and AI verification. Um so it's very interesting how uh blockchain with blockchain we started um from having like digital first NFTs that were like the uh notoriization of uh physical goods um and kind of like this the proper proof uh of existence and now we are coming back to a more grounding um like to the physical world which can be uh you know faces human faces AI retina palms and so on. Um so to conclude uh here I invite you also to look back at this

uh talk that we gave uh that I gave two years ago. uh is still on the Ethereum uh website the if Brussels um conference and I think you're going to see how the field actually has evolved so fast that we cannot just talk about the centralized AI um and the centralized um blockchain uh with the AI agents without considering AI safety and at the same time we need to uh consider um also proof of uh humanood and human verification as an important part of uh human in the loop for the process of of guaranteeing um cyber security and uh and safety in uh in most of transactions

and specifically agentic uh commerce. Thank you. >> Thank you Bianca. uh we have time for one question if there is uh any from the the public and in the meantime if panelist from the next panel can uh reach the stage if there isn't I do have one question actually you mentioned observability you mentioned data dog you mentioned a few other players um why do you think that's you know relevant and important to to add to your point to have a a good structure in observability >> yes so basically uh from what uh I'm

understanding the new protocols that have been developed uh by the TU foundation and in general on blockchain and crypto are trying to level up what for example this uh companies centralized you know like companies like Google and uh data dog are doing uh offchain observability basically means we have like a way to track whatever the agent is doing with time stamps um but this doesn't cure onchain so I think in a way what the Ethereum foundation is proposing is to recreate the same uh layer of structure on chain. However, this is not enough. So, uh having like the agent behavior uh completely like transparent is not enough to guarantee

uh that in the intent of the AI agent is aligned with the human intent. So that was my point. So observability uh it's not uh equivalent to uh interpretability which is let's say what happens in the brain of the AI which is the real black box. >> Got it. Got it. Absolutely. Now round of applause for Bianca.