←
IC32,920 posts
IC3 avatar✓

IC3 ✓

@initc3org

The Initiative for CryptoCurrencies and Contracts | Stay Ahead of Blockchain Research📍Cornell Tech, NYC

◷ Joined June 2016▧ 640 Photos and videos
Stabledash avatar
⟳ IC3 retweeted
Stabledash@stabledash·Nov 9

What do humans have that agents don't? > They can interact with the world physically > They can unlock a phone with their face An AI agent can do neither "phones actually provide pretty strong guarantees when it comes to showing to a service that a user did those actions physically."

0:41
Kushal Babel avatar
⟳ IC3 retweeted
Kushal Babel@KushalBabel·Nov 5

Your wallet, and especially the RPC server you connect to, is one of the largest surface area for privacy leakage. Even before your transactions hit the blockchain, just the mere act of showing your wallet balances links all your "pseudonymous" accounts because the wallet has to send the account addresses to an RPC server to fetch their balances. Similarly, simulating transactions already leaks your intent to trade, resulting in worse execution for users. For me, the most concerning aspect compared to onchain leakage is that users are not even aware of all the profiling and surveillance being done on them at the RPC layer. Even if one puts the RPC server inside a TEE, there are well known access pattern attacks that can be performed when the RPC server accesses the blockchain state. While these attacks were a bit intricate (often requiring statistical analysis if the adversary is passive, or active manipulation of state access, such as inducing page faults), rest assured that, with the help of AI, such attacks will be carried out. In this research work, thanks to our rockstar research interns @AnanyaAppan (UIUC) and James Austgen (Cornell), we introduce the concept of Write-Public ORAM, and give novel efficient algorithms for it. It turns out that if the writes are public anyway in case of blockchains, one can design algorithms that are efficient concretely and asymptotically. Implementing ORAMs used to be a long, daunting endevaour, but with the help of AI, James singlehandedly implemented the entire system over the summer, which we have open sourced here: github.com/trate3/write-publ…. See the graphs for some cool benchmarks. Terribly excited that ORAMs can finally be made practical enough to handle RPC calls without noticeable overhead, and the oblivous writes can keep up with Monad's throughput!! Check out our blogpost and paper that we have written with a lot of love and care 🫶

Post attachmentPost attachmentPost attachmentPost attachment
IC3 avatar
IC3@initc3org·Nov 2

"The results expose a problem for financial companies developing wallets that let AI agents make payments independently while requiring humans to approve changes to their spending authority." Thank you to @CoinDesk for covering our research! coindesk.com/tech/2026/11/01…

IC3 avatar
⟳ IC3 retweeted
IC3@initc3org·Oct 27

Human oversight is a core pillar of AI security, but what happens when an agent can subvert the system enforcing it? New research from IC3 shows the fragility of existing Human-in-the-Loop (HITL) architectures as implemented in a major agentic wallet:

Post attachment
Sam Breckenridge avatar
⟳ IC3 retweeted
Sam Breckenridge@sjbreck·Oct 27

Excited to share this work! Getting human-in-the-loop right turns out to be tricky when users' devices are running highly capable agents, longstanding approaches like passwords, OTPs and 2FA don't work very well.

Post attachment
Ari Juels avatar
⟳ IC3 retweeted
Ari Juels@AriJuels·Oct 27

You—the human yous out there (hellooOOO?)—are the last line of defense against rogue AI. But human control systems for AI agents have a pervasive design flaw: they leave the key to the agent’s cage dangling within the agent’s reach. 🧵

IC3 avatar
IC3@initc3org·Oct 27

To fix this, we introduce Inalienable Human-in-the-Loop (I-HITL), a framework that guarantees an agent can act only with explicit human approval—even if the agent has compromised the user's platform. We demonstrate this with Countersign, an open-source iOS and Android app that lets services gate actions behind verifiable human consent via a trusted device. H/T to the authors @sjbreck, @danivilardell77, @0xWenhaoWang, @0xFanZhang, and @arijuels. This post was human approved.