Three things in three different directions — a startup, a research project, and a social initiative in a developing country. They are my answer to one question: as agents simulate cognition and step into the physical world, where do humans stand? In working alongside them, are we the commander, the judge, the assisted — or fellow travelers who co-evolve and share the risk?
Final Year Project: an active-defense honeypot for LLM-agent attackers on ROS 2/DDS robotic systems, where a breach has real physical consequences.
Exploits the attacker’s own LLM weaknesses to fingerprint the intruding agent and elicit its objective — yielding threat intelligence on agent attackers.
Built a DDS-native, physics-aware decoy (Fast-DDS, rosbag2 replay, PyBullet) so an attacker interacts with it as if it were a real ROS 2 system.
CubeSieve: ML-Accelerated Cube Screening for Trivium Cube Attacks First author
FSE 2027 · Under review
Trained LightGBM models to rank promising cube candidates before formal cryptanalytic verification, cutting unnecessary algebraic-oracle checks while keeping final verification rigorous.
Tested generalization across rounds 600–660 and cube sizes 10–13; on a separate 80,000-cube benchmark, found 92% of maxterms by checking the top 60%, compared with 62% for random screening.
A Lightweight Block-Pair Image Steganography Method for UAV Imagery First author
IEEE ICSIPA 2026 · Published
Proposed a UAV image steganography scheme for covert data transmission over high-BER wireless channels.
Integrated BER-aware AODV routing with lightweight image steganography to improve robustness of hidden-data delivery under unstable UAV communication links.
Work
The Hong Kong University of Science and Technology (Guangzhou) Research Intern
AI Security
Designed and ran security experiments and tests on embodied AI systems, and wrote reports on the results.
Ran experiments for a PhD-led project that detects model substitution and quality degradation in third-party LLM API relay services.
Analyzed test data and wrote experiment reports that recorded test methods, results, and security findings.
Alibaba Cloud Campus Ambassador
Brand promotion and student community
Drove on-campus awareness for Alibaba Cloud products (cloud servers, AI tooling) through tech talks, community channels, and offline events.
Organized technical sharing sessions and hackathons connecting students with industry around cloud and AI.
Beijing LeadSec Network Information Technology Co., Ltd. Pre-sales Engineer Intern
Network security products
Delivered product demos and technical pitches; translated client requirements into customized network security solutions.
Owned pre-sales documentation and proposal materials; coordinated across pre-sales, engineering, sales, and clients to streamline deal cycles.
Endorsements
Dr. Li Jing
Assistant Professor · School of Computing and Data Science, XMU Malaysia
Yucheng combines unusually mature engineering taste with research-level rigor. He is one of the few undergraduates I have supervised who can identify a research-worthy gap on his own, then ship a working artifact toward it.
Prof. Xu Rongsheng (许榕生)
Senior Researcher & Doctoral Supervisor, Computing Center, IHEP, CAS · Chief Scientist, IHEP Network Security Lab · Recipient, National Science & Technology Progress Award (Special Class) · “Father of China’s Internet” · China’s first anti-hacker
Yucheng thinks innovatively; against agent attackers in robotic cybersecurity, he has proposed an interesting approach.
When choosing my undergraduate research direction, the field was wide open and no one was going to tell me which choice was "correct." I narrowed it through three layers of filtering. Trend — AI agents and embodied intelligence are durable directions, not fad waves, so I targeted robotic security. Infrastructure — the foundation had to be an open-source, hardware-decoupled ecosystem with long-term staying power, the way Linux became for operating systems; that pointed me to ROS 2. Whitespace — papers on agent-driven offensive security were exploding, but defensive work was nearly empty. I use LLMs daily and know they have "jagged intelligence" — strong in some dimensions, weak in others — meaning the attacker agent itself has exploitable weaknesses. A USENIX Security paper using Unicode honeytokens to counter LLM agents confirmed this direction was academically viable. These three layers converged into ROS2Pot — the first DDS-native, physics-aware honeypot for ROS 2 targeting LLM-agent attackers.
What's an original idea you've shipped?
Recomby.ai's AI Employee architecture. Most of the GEO (Generative Engine Optimization) market splits into two camps: pure SaaS tools (which leave all the work to the customer to figure out) or traditional agencies (which absorb all the work into human labor). I rejected both based on one core judgment: SEO optimizes keyword matching and can be fully automated; GEO optimizes the semantic representation of the business itself, which only works when "GEO experts" and "business experts" collaborate. So I designed the division of labor into two clean layers: we, as the GEO domain experts, only build skills that empower the agent — pipelines, domain prompts, professional tools; the customer, as the expert on their own business, just hands over the facts of their business to the system. The product form is a Feishu CLI (carrying data and scaffolding) + a Codex-class agent — the agent is what actually does the work. The result is a crushingly favorable cost structure for the customer: compared to hiring a full-time GEO employee, our offering is subscription-level overhead; compared to traditional agencies charging monthly for human labor, an agent doing the work brings the price down another tier. The customer doesn't need to become a GEO expert, doesn't need to maintain an in-house GEO team, and doesn't pay agency-level human markup — they just pay for skills + tokens.
How do you set priorities?
ROS2Pot is a large topic — protocol layer, physics layer, agent counteraction, attacker classification, ATT&CK mapping. Polishing every dimension to perfection before submitting to a top-tier venue would take 1.5 to 2 years and risk being scooped by peers. I made a deliberate trade-off: package the proven core — protocol-layer indistinguishability under professional ROS 2 scanning tools, physical consistency between the simulator and the real system, and a counteraction experiment that successfully elicited the attacker agent's operational objective — into a 3-page CCS poster submission first, and defer the rest to future work. This wasn't a compromise forced by limited capacity; it was a deliberately designed two-stage roadmap: claim the academic foothold during undergrad, then compound on it through grad school and PhD.
