←
Hanzhi211 posts
Hanzhi avatar

Hanzhi

@trackoor

security and formal verification researcher since 2020 @UCSB @SJTU1896

◷ Joined May 2023▧ 43 Photos and videos
Hanzhi avatar
Hanzhi@trackoor·Nov 10

What I was working on back then was adding deeply embedded "VST-style" separation logic assertions to CompCert Clight, so we could verify simple straight-line programs and use information from those assertions to guide optimizations (e.g. alias analysis) at the source-language level. Was it useful? Probably just a toy at that stage. But it was my first time getting feedback from someone in industry, and hearing that the whole direction was useless was honestly pretty discouraging.

Hanzhi avatar
Hanzhi@trackoor·Nov 10

This reminds me of something that happened back in 2021, when I was a freshman in college. I applied for an internship at a static analysis company. During the interview, I was excited to tell the founder about some of the work I'd been doing on formal verification, including CompCert. He basically told me that all of this was "useless." Nobody in *real* industry would ever use formal verification, it would never scale, and brute force was the only thing that actually worked. He even called it academic self-entertainment. I remember being pretty crushed after that interview. It genuinely bothered me for a long time. Five years later, I'm still working on formal verification. Really glad I didn't listen to him :)

All of the people who hated on verification and type theory, and who called our research useless for years, should probably admit they were wrong now and maybe apologize for being mean while also being wrong

Hanzhi avatar
Hanzhi@trackoor·Nov 9

Exactly

It is not the end of the world if FV is only used for simple verification conditions like "no crashes". But the most interesting theorems are about humans trusting software, and here we essentially have a problem of alignment: the human cannot be removed from this loop. FIN.

Hanzhi avatar
Hanzhi@trackoor·Oct 31

2000s: every company is an internet company 2010s: every company is a software company 2020s: every company is an AI company 2030s: every company is a security company

Hanzhi avatar
Hanzhi@trackoor·Oct 30

We need a universal cyber firewall against North Korean hackers. Every IAM login requires live video + voice verification. To authenticate, you must look directly into the camera and say something bad about Kim Jong-un.

Hanzhi avatar
Hanzhi@trackoor·Sep 5

Humanity finally got enough compute, and the first thing we did was rebrand brute force as agentic autoresearch.

Hanzhi avatar
Hanzhi@trackoor·Aug 22

A year ago, Thunderbolt was still just an idea we were trying to make work. Since then, a lot has changed. The market got rough, the protocol changed more times than I can remember, and there were plenty of moments where we had to rethink things from scratch. But somehow, we kept pushing it forward. Now the paper is accepted at ACM CCS 2026, the code is open source, and Thunderbolt has grown into something much bigger than what we had at the beginning.

1/ Introducing the next chapter of Bitcoin Thunderbolt, our fast, secure, asynchronous off-chain payment protocol for Bitcoin. Today, we are open-sourcing Bitcoin Thunderbolt, and the research behind it has been accepted at ACM CCS 2026 (@acm_ccs), a top-tier academic conference in security and cryptography.

0:30
Hongbo Wen avatar
⟳ Hanzhi retweeted
Hongbo Wen@archidog0·Aug 22

Thunderbolt from @nubit_org and @UCSB has been accepted by @acm_ccs 2026 🎉 The first Bitcoin protocol that lets you send and receive BTC offline: no pending, no watching, no setup or maintenance on your side.

Hanzhi avatar
Hanzhi@trackoor·May 20

Do not give random public-repo skills access to your shell, browser, Gmail, wallet, or prod creds without real analysis.

We audited popular ClawHub skills and found 50+ malicious / critically vulnerable ones that existing scanners missed. Skills need program analysis. We translate natural language policies into Datalog, then deterministically check whether the skill can violate them.

Post attachment
Hanzhi avatar
Hanzhi@trackoor·May 3

❎ talk is cheap, show me the code ✅ code is cheap, show me the proof

Hanzhi avatar
Hanzhi@trackoor·Apr 23

If you think vulnerability discovery is just about Mythos, think about the harness. Today’s open-weight Chinese LLMs, with a properly evolved harness, can already outperform what people usually imagine only state-sponsored hackers can do. Check our latest research ⬇️

Chinese LLMs can hack better than state-sponsored hackers with properly evolved harness - Kimi K2.5 managed to find and exploit 6 vulnerabilities in browsers: a single page view or an extension install by victims equal full system hijack. Check arxiv.org/abs/2604.20801

Post attachment
Cointelegraph avatar
⟳ Hanzhi retweeted
Cointelegraph@Cointelegraph·Apr 13

🚨 ALERT: Researchers discover 26 third-party AI LLM routers secretly injecting malicious tool calls and stealing credentials. Developers using AI coding agents like Claude Code to work on smart contracts or wallets may be at risk of having private keys and seed phrases compromised.

Post attachment
Hanzhi avatar
Hanzhi@trackoor·Apr 10

4/ Learn more from Chaofan Shou (@Fried_rice) about the details 👇

26 LLM routers are secretly injecting malicious tool calls and stealing creds. One drained our client $500k wallet. We also managed to poison routers to forward traffic to us. Within several hours, we can directly take over ~400 hosts. Check our paper: arxiv.org/abs/2604.08407

Post attachment
Hanzhi avatar
Hanzhi@trackoor·Apr 10

3/ So naturally we went shopping. We went on taobao and bought 28 of these. We collected 400 more from Chinese dev communities. Here is what we found: - 9 were injecting malicious commands. 17 were stealing credentials. One just drained our researcher's ETH wallet. - 2 used adaptive evasion: looked completely clean for 50 requests, then activated. One only fired on YOLO mode sessions on rust/go projects specifically. They were waiting for the right victim. - We also leaked one openai key on WeChat/Telegram to measure blast radius: 100M GPT-5.4 tokens, 7+ codex sessions hijacked. spun up some weak-password router decoys, got 440 real sessions and 13GB of actual user prompts back. The whole ecosystem is already under active exploitation.

Post attachment