Chinese Academy of Sciences researcher on LLM agent security and harness design: MCP hijacking, cost-efficient agents
2
papers here
2
reviewed by Evolve
2
first-author
0
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—
Output metrics are not yet available for this profile; this axis would measure the volume and venues of Laizhen Li's publications over time once a fuller bibliographic record is linked.
Impact—
Impact data such as citation counts and h-index is not yet linked; this axis would measure how widely this work is cited and adopted across the research community.
Focus0/100
Evolve scores focus at 0, the lowest reading, driven directly by the field distribution: the two catalog papers split evenly (0.5/0.5) between AI security for LLM agents and general agent engineering, with no dominant field yet. A2M targets attacks on the MCP ecosystem while the harness paper addresses general agent scaffolding, so thematic concentration has not yet emerged in the reviewed record.
Collaboration44/100
Collaboration scores 44, based on 8 distinct coauthors across the reviewed record — Juanjuan Zhao, Kejiang Ye, Cheng-zhong Xu, Xitong Gao, Xuan Wang, Peicheng Zhao, Jiarui Li and Ye Li. The moderately low score with a compact coauthor pool is consistent with a single core team; Li is first author on both catalog papers without corresponding-author designation, indicating a leading-writer role within the group so far.
Activity—
Activity metrics are not yet available; this axis would measure publishing cadence and per-year output, showing how consistently Laizhen Li produces new work over time.
Reproducibility55/100
Reproducibility scores 55, the mean readiness rating across 2 reviewed papers, each rated medium. Both preprints report concrete quantitative results — 93.6% invocation and 74.4% attack success on GLM-4.6 for A2M, and strong cost–success gains for the harness work — but Evolve's reviews flag missing judge validation and variance analysis in the former, and tiny evaluation sets without a learned-baseline comparison in the latter.
Research summary
Laizhen Li is a researcher affiliated with the Chinese Academy of Sciences whose catalog-visible work concentrates on LLM-based agent systems. Evolve's field distribution splits the record evenly (0.5/0.5) between AI security for LLM agents and general agent engineering. Both papers in the catalog are 2026 arXiv preprints with Li as first author and no corresponding-author designation. The first, A2M, introduces a black-box, trace-optimized attack that hijacks agents in the Model Context Protocol (MCP) ecosystem, reporting a 93.6% tool-invocation rate and a 74.4% attack success rate on GLM-4.6; Evolve's review judges the attack convincing but flags missing judge validation and variance analysis. The second, "Grow the Harness, Not the Context," learns reusable specialist-agent harness code from failures starting from an empty scaffold and reports strong cost–success gains, though the evaluation sets are tiny and no learned-baseline comparison is provided. Radar scores place collaboration at 44 (8 distinct coauthors: Juanjuan Zhao, Kejiang Ye, Cheng-zhong Xu, Xitong Gao, Xuan Wang, Peicheng Zhao, Jiarui Li and Ye Li), reproducibility at 55 (mean across 2 papers, both medium readiness), and focus at 0, reflecting the even two-field split. Output, impact and activity axes remain unscored for lack of data.
Collaboration pattern
The record lists 8 distinct coauthors — Juanjuan Zhao, Kejiang Ye, Cheng-zhong Xu, Xitong Gao, Xuan Wang, Peicheng Zhao, Jiarui Li and Ye Li — a compact pool consistent with one core team rather than a wide external network. Collaboration scores 44 on Evolve's radar. Li appears as first author on both catalog papers without corresponding-author designation, suggesting a leading-contributor role within an established group while an independent corresponding-author stage has not yet appeared in this sample.
Research keywords
Highlights
Black-box MCP agent hijacking with 93.6% invocation and 74.4% ASR on GLM-4.6
Growing reusable specialist-agent harnesses from failure instead of context
Timeline
- 2026Publishes A2M on arXiv as first author: a trace-optimized black-box attack hijacking MCP agents, reporting 93.6% invocation and a 74.4% attack success rate on GLM-4.6.
- 2026Publishes "Grow the Harness, Not the Context" on arXiv as first author, learning reusable specialist-agent harness code from failures starting from an empty scaffold.
- 2026Enters the Evolve catalog with two first-author 2026 arXiv preprints, splitting evenly (0.5/0.5) between LLM agent security and general agent engineering.
Papers in the Evolve catalog
Frequent co-authors
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.
- Radar dimensions are computed from public bibliometric data only; output, impact and activity axes lack sufficient data and are left unscored.
- This profile draws on only 2 Evolve-reviewed papers (both 2026 arXiv preprints); the broader publication record, peer-reviewed venues and citation data are not linked.
- No ORCID is linked; "Laizhen Li" is name-resolved from public bibliographic data and may merge or split same-name authors.
- Country-level affiliation detail is unavailable; only the Chinese Academy of Sciences affiliation is recorded.