Early-career LLM agent security researcher at Northern Arizona University: prompt-injection benchmarks and detectors
3
papers here
3
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
This axis would measure publication volume and its growth over time. It is not scored yet: the snapshot records only three 2026 arXiv works, with h-index and i10-index both at 0, which is too short a history to characterize research output.
Impact—
Impact would capture citation uptake and scholarly influence. No citation evidence exists in this snapshot — h-index and i10-index are both 0 and no citation timeline is recorded — so the axis is left unscored rather than estimated.
Focus0/100
The focus axis scores 0, and the raw field distribution explains it: output splits evenly three ways — prompt-injection detection (0.33), prompt-injection defense (0.33), and LLM-based multi-robot systems (0.33). Two of the three labels still sit inside LLM agent security, so the work is thematically coherent even though no single field yet dominates.
Collaboration28/100
Collaboration scores 28, matching a compact network: the OpenAlex snapshot lists only 3 distinct co-authors, and the recurring names are Mithun Paul Saint-Germain, Chongwen Dong, and Carlo R. daCunha. Both first-authored security papers and the middle-position robotics paper circulate within this small team, suggesting a tight working group rather than a broad cross-institutional web.
Activity40/100
Activity scores 40 on a very short record: recent-3-year works number 3, and the per-year counts list 2026 alone, with 3 works. All catalog entries are 2026 arXiv preprints, so output concentrates in a single year and the trajectory cannot yet be read as ramping up or plateauing.
Reproducibility70/100
Reproducibility is the strongest axis at a mean score of 70 across 3 assessed papers. Evolve rates DriftNet 'high' for repro readiness; the AgentDrift benchmark and the robotics paper are both 'medium', the former because only a surface baseline probes benchmark difficulty and the latter because results come from single-run, seedless simulations.
Research summary
Asif Pinjari is an early-stage researcher at Northern Arizona University (US) whose 2026 output centers on the security of LLM-based agents, with a secondary line in LLM-guided multi-robot systems. The Evolve catalog holds three arXiv preprints, all from 2026, and Pinjari is first author on two of them. DriftNet proposes a dual-head trajectory transformer that both issues an injection verdict and labels the hijacked step, reporting F1 0.983 — Evolve calls the results rigorous but notes they are self-benchmarked on synthetic, single-generator data. Its companion, AgentDrift, is positioned as the first public step-level injection-labeled benchmark of agent trajectories; the review credits it with unusual honesty about its flaws while observing that only a surface baseline tests its difficulty. The third paper, a middle-author contribution, combines a schema-bounded LLM with UCB-guided policy refinement and local Double DQN for decentralized multi-robot navigation, reaching a best median of 42 ticks in configuration C4 under single-run, seedless simulation. Bibliometrics reflect a nascent record: h-index and i10-index are both 0, recent-3-year works number 3, and only 3 distinct co-authors appear (recurring: Mithun Paul Saint-Germain, Chongwen Dong, Carlo R. daCunha). Radar scores place reproducibility as the strongest axis (mean 70 across 3 papers), with activity at 40, collaboration at 28, and focus at 0 given the even three-way field split. Output and impact are not yet scored.
Collaboration pattern
Pinjari's collaboration footprint is small and stable: OpenAlex records 3 distinct co-authors, and the recurring names are Mithun Paul Saint-Germain, Chongwen Dong, and Carlo R. daCunha. The pattern suggests a tight core team producing both the agent-security line and the multi-robot line together, with no cross-institutional co-authorship visible in this snapshot.
Research keywords
Highlights
Dual-head trajectory transformer detects and localizes prompt injection
Step-labeled benchmark of injection-hijacked agent trajectories
Schema-bounded LLM refinement for decentralized multi-robot policies
Timeline
- 2026Co-authors a schema-bounded LLM study on decentralized multi-robot navigation (middle author position).
- 2026Releases AgentDrift, a step-labeled benchmark of injection-hijacked LLM agent trajectories, as first author.
- 2026Publishes DriftNet, a dual-head trajectory transformer for detecting and localizing prompt injection, as first author.
Papers in the Evolve catalog
- first author
DriftNet: A Dual-Head Trajectory Transformer for Detecting and Localizing Prompt Injection in LLM Agents
- first author
AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories
A Schema Bounded Language Model for Refining Robot Policies Without Destabilizing Local Learning
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 counts (h-index, i10-index, co-author and per-year counts) come from an OpenAlex snapshot; the profile is name-resolved from public bibliographic data and no ORCID was provided.
- All cataloged works are 2026 arXiv preprints; no peer-reviewed venues, citation data, or career history fall within this data scope.
- Output and impact radar axes are unscored because the underlying data is insufficient; paper quality notes reflect Evolve's own reviews, not external peer review.
- Headline numbers (F1 0.983, 42 ticks median) come from the authors' own setups — synthetic single-generator data and single-run seedless simulation — so they are reported results, not independently verified.