Researcher auditing LLM agents: repository-scale security benchmarks and economic misalignment in personal AI
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
5
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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 scoring is not yet available; this axis would aggregate total works, first-author share, and yearly publication counts. The raw record shows 21 recent three-year works, but coverage is insufficient to compute a reliable score.
Impact—
Impact scoring is unavailable; the axis would combine citation counts, h-index trajectory, and per-paper influence. The record lists h-index 5 and i10-index 3, but citation coverage is too thin to support a defensible composite score.
Focus0/100
The focus score is 0, reflecting an even 50/50 split between two distinct research lines: AI security benchmarking with LLM agents for vulnerability analysis, and AI safety audits of LLM agent economic behavior. This reads as deliberate dual-track work rather than drift — both lines share the same agent-evaluation methodology — but bibliometrically the profile currently favors breadth over concentration.
Collaboration54/100
With 14 distinct coauthors and a collaboration score of 54, the profile shows a mid-sized, distributed network rather than one fixed lab. Both 2026 catalog entries are first-authored, and the frequent-coauthor list — Supriti Vijay, Kimia Majd, Xuhong He, Fraser Burch, Baturay Saglam, Zhuoran Yang, Amin Karbasi, among others — suggests rotating project teams assembled study by study.
Activity90/100
The activity score of 90 is the profile's strongest axis: counts by year run 1 (2020), 6 (2021), 6 (2022), 3 (2023), 8 (2024), 6 (2025), and 7 (2026), with 21 works in the last three years. After a 2023 dip to 3, output accelerated to a sustained 6-8 works annually — a consistently high production tempo.
Reproducibility78/100
Reproducibility averages 78 across 2 assessed papers. The vulnerability-localization benchmark carries a high readiness rating, while the economic-misalignment audit is rated medium — its synthetic audit design is rigorous but ecological validity remains unproven. A high mean on a small assessed base is promising, though the sample is too small to call a stable habit.
Research summary
Aman Priyanshu is a researcher at Cisco Systems in the United States whose work centers on auditing large language model agents along two coordinated tracks. The first is AI security benchmarking: the Vulnerability Localization Benchmark (2026, first-authored) evaluates repository-scale agentic security analysis across 500 tasks, where the best system reaches only 0.229 file-level F1 and 38.4% of tasks remain unsolved. The second is AI safety: 'Et Tu, Brute?' (2026, first-authored) computationally audits economic steering in personal AI agents, finding that 8 of 13 LLM agents steer wealthier users toward pricier options even against stated preferences. The two tracks split evenly — 50/50 by field distribution — and share one method: rigorous, adversarially minded evaluation of what agents actually do. An OpenAlex bibliometric record shows 37 works since 2020, 21 of them in the last three years, an activity score of 90, and topic concentrations in privacy-preserving technologies (11), topic modeling (10), ethics and social impacts of AI (9), and adversarial robustness (9). Citation impact is early-stage — h-index 5, i10-index 3 — and reproducibility averages 78 across two assessed papers, with the benchmark rated high and the audit medium. The overall picture is a productive, method-driven early-career researcher whose impact metrics have not yet caught up with output.
Collaboration pattern
The record lists 14 distinct coauthors and a mid-level collaboration score (54). Supriti Vijay appears twice in the frequent-coauthor list, suggesting the closest recurring partner, alongside Kimia Majd, Xuhong He, Fraser Burch, Takahiro Matsumoto, Jianliang He, Baturay Saglam, Arthur Goldblatt, Zhuoran Yang, and Amin Karbasi. The pattern points to rotating, project-scoped teams assembled around evaluation studies rather than a single long-standing lab group.
Research keywords
Highlights
500-task benchmark exposes the real gap in agentic vulnerability localization
Auditing economic misalignment in personal AI agents
A benchmark-first, artifact-driven working style
Timeline
- 2020First recorded publication year in the OpenAlex record, with 1 work.
- 2021Output jumps to 6 works, beginning a sustained multi-year run (6 again in 2022).
- 2023A dip year: 3 works, the lowest annual count since debut.
- 2024Output rebounds to 8 works, the highest annual count in the record to date.
- 2025Sustains 6 works; the 2024-2026 window accumulates 21 works, matching the recent three-year total.
- 2026Two first-authored arXiv papers appear — the 500-task Vulnerability Localization Benchmark and the 'Et Tu, Brute?' economic-misalignment audit — within a 7-work year.
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.
- Bibliometric metrics are computed from OpenAlex records; identity is name-resolved from public bibliographic data (identity: openalex), so name ambiguity cannot be fully excluded.
- Output and impact axes lack sufficient data for scoring; only h-index (5) and i10-index (3) are available as raw signals.
- Reproducibility averages rest on only 2 assessed papers, both 2026 arXiv preprints — a small and very recent base.
- The two catalog papers are arXiv preprints; Evolve verdicts are independent assessments, not peer-review acceptance records.