Cheng-zhong Xu

University of Macau

Unclaimed · AI-structured profileOpenAlex

University of Macau researcher in LLM agent security and agent harness design, profiled from two 2026 arXiv papers

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-02) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Not yet scored: the snapshot contains no publication-volume or productivity metrics for this researcher. This axis would measure total output, its distribution over years, and venue spread once a fuller bibliographic record is linked.

Impact—

No citation, h-index, or reference data is available in this snapshot, so impact is not scored. The axis would measure citation uptake, influential works, and how widely the reviewed 2026 papers are being used.

Focus0/100

The focus score is 0, and the field distribution explains why: the reviewed catalog splits exactly 50/50 between AI security (LLM agent security) and general work. One 2026 paper attacks the MCP agent ecosystem, the other builds agent harnesses from an empty scaffold, so current evidence shows a two-sided agenda around LLM agents rather than a single concentrated niche. More cataloged output would be needed to confirm whether the split is strategic breadth or an emerging focus.

Collaboration44/100

With a collaboration score of 44 and 8 distinct coauthors in the snapshot, this researcher works in modest-size teams rather than sprawling consortia. Both reviewed 2026 arXiv papers list the researcher in a middle author position without corresponding-author role, and recurring partners include Laizhen Li, Juanjuan Zhao, Kejiang Ye, Xitong Gao, Xuan Wang, Peicheng Zhao, Jiarui Li, and Ye Li — a compact co-authorship circle around the reviewed 2026 work.

Activity—

Insufficient data: only two reviewed arXiv papers, both dated 2026, cannot establish a publication cadence. This axis would measure recent yearly output and research tempo once a longer publication history is attached.

Reproducibility55/100

The reproducibility score of 55 is a mean taken over 2 reviewed papers, both rated medium reproducibility readiness. Evolve's verdicts flag the concrete gaps: the A2M attack reports 93.6% invocation and 74.4% attack success on GLM-4.6 but lacks judge validation and variance analysis, while the harness-learning paper shows strong cost–success gains yet uses tiny evaluation sets with no learned-baseline comparison. Methods look implementable, but reported claims still need independent verification.

Research summary

Cheng-zhong Xu is a researcher at the University of Macau whose current Evolve footprint centers on large language model (LLM) agents — both attacking and building them. The catalog contains two 2026 arXiv papers, split exactly 50/50 between AI security (LLM agent security) and general agent tooling. The first, A2M, presents a black-box attack that hijacks agents in the Model Context Protocol ecosystem, reaching 93.6% invocation and a 74.4% attack success rate on GLM-4.6 through trace-optimized adversarial optimization. The second, "Grow the Harness, Not the Context," learns agent harness code from failures starting from an empty, strategy-free scaffold, turning failures into reusable specialist agents with strong cost–success gains. On both papers the researcher appears in a middle author position without corresponding-author status, working within a compact circle of 8 distinct coauthors that includes Laizhen Li, Juanjuan Zhao, Kejiang Ye, Xitong Gao, Xuan Wang, Peicheng Zhao, Jiarui Li, and Ye Li. The reproducibility score of 55 — a mean over these 2 papers, both rated medium readiness — reflects promising but not yet independently verified results. Output, impact, and activity metrics are unavailable in the snapshot, so this is an early-stage profile: name-resolved from public bibliographic data and limited to reviewed 2026 arXiv work.

Collaboration pattern

The snapshot records 8 distinct coauthors and a collaboration score of 44, pointing to compact teams rather than large consortia. On both 2026 papers the researcher sits in a middle author position without corresponding-author role, suggesting contribution within a led-by-others structure. Recurring names include Laizhen Li, Juanjuan Zhao, Kejiang Ye, Xitong Gao, Xuan Wang, Peicheng Zhao, Jiarui Li, and Ye Li, forming a stable co-authorship circle around the MCP-security and agent-engineering work.

Research keywords

Model Context Protocol (MCP)LLM agentsPrompt injectionTool selection attackAdversarial optimizationSupply-chain securityBlack-box attackAgent harnesses and scaffoldsCost–success trade-offsReusable specialist agents

Highlights

A2M: black-box hijacking of MCP agents with trace-optimized attacks
The paper proposes A2M, a black-box attack that hijacks agents in the Model Context Protocol ecosystem, reporting 93.6% invocation and a 74.4% attack success rate on GLM-4.6. Evolve finds the attack convincing while noting missing judge validation and variance analysis.
Growing reusable specialist agents from an empty scaffold
The "Grow the Harness, Not the Context" paper learns agent harness code from failures starting from a strategy-free scaffold, reporting strong cost–success gains; Evolve flags tiny evaluation sets and the absence of a learned-baseline comparison as open questions.
A two-sided 2026 agenda: attacking and engineering the agent stack
Both 2026 arXiv entries split evenly (50/50 in the field distribution) between LLM agent security and general agent tooling, sketching a researcher working simultaneously on the attack and engineering sides of the MCP-era agent ecosystem.

Timeline

2026The A2M paper appears on arXiv, introducing a trace-optimized black-box attack that hijacks MCP agents, with 93.6% invocation and a 74.4% attack success rate reported on GLM-4.6.
2026The second arXiv entry, "Grow the Harness, Not the Context," learns agent harness code from failures on a strategy-free scaffold, reporting strong cost–success gains.
2026The two papers define the current Evolve footprint: an even 50/50 split between LLM agent security and general work, 8 distinct coauthors, and a mean reproducibility score of 55.

Papers in the Evolve catalog

Frequent co-authors

Laizhen Li×2 Juanjuan Zhao×2 Kejiang Ye×2 Xitong Gao×2 Xuan Wang×1 Peicheng Zhao×1 Jiarui Li×1 Ye Li×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometrics in this profile come only from an OpenAlex-derived Evolve snapshot covering 2 reviewed arXiv papers, both dated 2026.
No ORCID or country is attached and the name may be common; identity is name-resolved from public bibliographic data, not independently confirmed.
Output, impact, and activity dimensions are unscored because citation counts, publication volume, and multi-year history are absent from the input data.
The researcher is a middle author on both papers and corresponding author on none, so leadership and career-stage signals are weak.