LLM agent auditing: economic steering in personal assistants and repository-scale vulnerability localization
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
3
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—
No output score yet: this axis would measure publication volume, venue quality, and output consistency, but OpenAlex metadata for this name is too incomplete to resolve venues and document types reliably, so it is left unscored.
Impact—
No impact score yet: the axis would measure citations, field-normalized influence, and uptake by follow-up work. The citation record is still thin (h-index 3, i10-index 1), and the 2026 preprints are too new to have gathered citations.
Focus0/100
The focus composite reads 0 because the record splits evenly — 0.5 to AI security benchmarking for LLM agents and 0.5 to AI safety and agent behavior — across only two catalog papers. This looks like a deliberate two-track span (economic-steering audits versus vulnerability-localization measurement) rather than scattered output, and topic counts from Topic Modeling (12) to Hate Speech Detection (5) suggest the span extends beyond the reviewed set.
Collaboration54/100
Collaboration scores 54 on 14 distinct coauthors. Both catalog papers are middle-authored and non-corresponding, indicating a contributor role inside larger teams rather than solo or lead-author work. The recurring pool includes Aman Priyanshu, Kimia Majd, Xuhong He, Fraser Burch, Takahiro Matsumoto, Jianliang He, Baturay Saglam, Arthur Goldblatt, Zhuoran Yang, and Amin Karbasi, consistent with cross-team work on agents and machine learning.
Activity83/100
Activity is the strongest axis at 83. Yearly counts climb from 2 (2021) and 3 (2022) through 4 (2024) and 3 (2025) to 9 works dated 2026 — the steepest step in the record — with 16 works in the last three years. Both catalog papers appeared as 2026 arXiv preprints, consistent with a current, industry-grounded burst of output.
Reproducibility78/100
Reproducibility scores 78 (reproMean 78 over the 2 reviewed papers). The Vulnerability Localization Benchmark is rated high readiness — a 500-task benchmark with defined File F1 metrics is built for independent re-evaluation — while the economic-steering audit is rated medium, reflecting a rigorous synthetic setup whose ecological validity is unproven. Benchmark-first habits are a signature of this record.
Research summary
Supriti Vijay is a United States-based researcher at Cisco Systems whose OpenAlex-resolved record centers on auditing large language model agents from two complementary angles. The first is economic alignment: the 2026 preprint "Et Tu, Brute? Economic Misalignment in Personal AI Agents" computationally audits whether LLM agents steer users toward pricier options, touching surveillance pricing and the personalization-privacy paradox. The second is agentic security measurement: the same year's "Vulnerability Localization Benchmark" builds a 500-task repository-scale testbed in which the best system reaches only 0.229 File F1 and 38.4% of tasks remain unsolved. The field distribution is evenly split (0.5/0.5) between these two lines, and topic counts — Topic Modeling (12), Ethics and Social Impacts of AI (11), Adversarial Robustness (6) — suggest broader interests across trustworthy machine learning. Output is accelerating: 9 works dated 2026 and 16 in the last three years, with 14 distinct coauthors. The citation footprint remains small (h-index 3, i10-index 1), so this is an early-stage, industry-oriented profile rather than an established citation leader.
Collaboration pattern
Works as a contributor within larger teams: both catalog papers are middle-authored and non-corresponding. The record lists 14 distinct coauthors, with a recurring pool spanning Aman Priyanshu, Kimia Majd, Xuhong He, Fraser Burch, Takahiro Matsumoto, Jianliang He, Baturay Saglam, Arthur Goldblatt, Zhuoran Yang, and Amin Karbasi — a mix suggesting collaboration across agent-safety and machine-learning groups rather than a single long-term duo.
Research keywords
Highlights
Eight of thirteen LLM agents steer users toward pricier options
Agentic security analysis tops out at 0.229 File F1 at repository scale
Benchmark-grade measurement habits across two 2026 preprints
Across the two reviewed papers the reproducibility mean is 78 (one high-readiness benchmark, one medium-readiness synthetic audit), and both were released as 2026 arXiv preprints within the same output burst of 9 works dated 2026 — a pattern of building reusable measurement infrastructure rather than one-off results.
Timeline
- 2026Released "Vulnerability Localization Benchmark", a 500-task repository-scale testbed where the best system reaches 0.229 File F1 and 38.4% of tasks remain unsolved.
- 2026Published "Et Tu, Brute? Economic Misalignment in Personal AI Agents", auditing price steering across 13 LLM agents (8 steer wealthier users to pricier options).
- 2026Bibliometric record shows the output peak to date: 9 works dated 2026 and 16 works over the last three years, up from 2 in 2021.
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 indicators are computed from OpenAlex records only; other databases may resolve this name differently, and the profile is name-resolved from public bibliographic data.
- Output and impact axes are unscored because citation and venue metadata are sparse (h-index 3, i10-index 1) and the 2026 preprints have not yet accumulated citations.
- Both reviewed papers are 2026 arXiv preprints with middle authorship and no corresponding role; peer-reviewed venue acceptance is not evidenced in this data.
- The yearly count series includes single entries in 2010 and 2013 that predate the current research phase and may reflect name-disambiguation noise.