Xiaohang Yu

~/sky/research/directions

• Agent security: cryptographic protocols for agent credential isolation and secret-use delegation (SUDP, first author).
• Agentic forensics: agentic frameworks for blockchain forensics (LOCARD, PhD work, IEEE ICBC 2026).
• Reinforcement learning: RL post-training for LLM reasoning (ACL 2026 main conference; TMLR 2025 agentic-RL survey, co-first author).

~/sky/research/field-map

~/sky/research/publications

• SUDP: Secret-Use Delegation Protocol for Agentic Systems code arXiv preprint, 2026 · first author
• LOCARD: An Agentic Framework for Blockchain Forensics code IEEE ICBC 2026 · first author
• Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning ACL 2026 (main)
• The Landscape of Agentic Reinforcement Learning for LLMs: A Survey paper list TMLR 2025 · co-first author
• earlier
• The Design of a Visual Tool for the Quick Customization of Virtual Characters in OSSL Yungang Wei, Xiaoran Qin, Xiaoye Tan, Xiaohang Yu, Bo Sun, Xiaoming Zhu · IEEE Cyberworlds 2015, pp. 314 to 320
• Exploring the Use of a 3D Virtual Environment in Chinese Cultural Transmission Yungang Wei, Xiaoye Tan, Xiaoran Qin, Xiaohang Yu, Bo Sun, Xiaoming Zhu · IEEE Cyberworlds 2014, pp. 345 to 350

~/sky/projects/safeclaw

An agent that holds your API key holds your full authority. Every call, forever.

SafeClaw ends that. Agents only ever see phantom credentials; a local proxy injects the real secret at the network edge, and sensitive calls wait for a passkey tap. Each tap is a single-use grant bound to one exact operation, so even a fully compromised agent cannot spend your authority.

co-founded 2026 · production stack built solo in three months · open source

safeclaw.pro github the paper behind it

~/sky/projects/open-source

• Awesome-AgenticLLM-RL-Paperspaper list companion to the TMLR 2025 agentic-RL survey; started and maintained here, later handed to a co-authormarkdown · 1898 stars
• tg-claude-botTelegram bridge for Claude Code: resume any CLI session, per-topic chats, local voice, inline-button approvalsPython · 16 stars
• ida-evm-enhancedIDA processor module for the EVM, extended for real-world contract reversingPython · 7 stars
• locardfirst agentic framework for blockchain forensics: LLM agents tracing transactions across chainsPython · 5 stars
• sudpSecret-Use Delegation Protocol, the protocol layer under SafeClawRust · 4 stars
• rwctf2018finals-AcoraidaMonicaReal World CTF 2018 Finals blockchain challenge, the first EVM jump-oriented-programming puzzleJavaScript · 4 stars
• defcon27slides for the EVM Opcode JOP talk, DEF CON 27 Blockchain Villageslides · 3 stars
• masframework for modular autonomous multi-agent systems, execution flow separated from the task graphPython · 1 star
• opmloptimistic machine learning: off-chain inference verified on-chain through a fault-proof dispute engine; core technical contributor to the standardSolidity
• ctf-eth-envdockerized private Ethereum network with an RPC proxy, for CTF challenge infrastructureJavaScript
• chainkitSolidity helper for deploying contracts from raw bytecode and reading deployed codeSolidity

ordered by stars as of september 2026

~/sky/projects/standards

• Core technical contributor to open standards: opML (optimistic machine learning), ERC-7007 (onchain AI-generated assets), ERC-7641 (fee-sharing token).
• Technical lead of China’s first national blockchain vulnerability scoring standard (CNVD); participated in 3 MIIT national industry-standard research projects.

~/sky/talks/selected

• IEEE ICBC 2026 LOCARD: An Agentic Framework for Blockchain Forensics Brisbane · upcoming
• DEF CON 27 Village EVM Opcode JOP Las Vegas · 2019
• DEF CON China 1.0 Village Practical Security of Blockchain in Industry Beijing · 2019
• ISC Tech Forum Guide to Byzantine Traitors: New Types of Blockchain Vulnerabilities Beijing · 2020
• Real World CTF Tech Forum Defeating Byzantium: Blockchain Security in Practice Beijing · 2019
• Yunqi 2050 Smart Contract Offense and Defense Hangzhou · 2018

six of nine

~/sky/about/career

• since 2026 Co-founder, SafeClaw: an open-source credential broker for AI agents, so agents get scoped, per-use, user-approved delegation instead of reusable secrets. Built the production stack solo: a Rust egress proxy, an end-to-end encrypted passkey-sealed vault, a web console, and a cloud control plane.
• 2023 to 2025 Technical co-founder of a blockchain infrastructure company. Built and led an engineering team of 10+; research and implementation of core protocols for verifiable AI systems and zero-knowledge-proof infrastructure.
• 2018 to 2023 Head of blockchain security, Chaitin Tech. Founded and led the practice from zero, one of the earliest such teams in China. Around 100 zero-day vulnerabilities across the consensus, P2P, and verification layers of major blockchains; roughly 1M lines of code audited; 5 CVEs from smart-contract audits.

~/sky/about/education

• Ph.D. Computer Science, Imperial College London, 2021 to 2027 (expected). Advised by William Knottenbelt.
• M.S. Security Informatics, Johns Hopkins University. China Scholarship Council national scholarship.
• B.S. Computer Science, Beijing Normal University.

~/sky/about/honors

• DEF CON CTF World Finals, 3rd place (team Tea Deliverers)
• Qiangwang Cup, national cybersecurity competition, first prize, twice
• ACM-ICPC Asia Regional, silver medal, twice
• China Undergraduate Mathematical Contest in Modeling, national first prize
• OSCP · PMP