Sabrina Sadiekh

—

Unclaimed · AI-structured profileORCID OpenAlex

LLM security researcher: adversarial agent benchmarks and compact guardrails against prompt injection and jailbreaks

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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 lacks publication-volume data beyond two 2026 arXiv preprints. This axis measures total research output — number, type, and consistency of publications across years and venues — and needs a fuller bibliographic record before it can be assessed.

Impact—

Not yet scored: citation counts, h-index, and downstream uptake are unavailable in the input data. This axis would measure scholarly influence — citations, adoption of the DUMA-Bench benchmark and HiveTraceGuard-Pro guardrail, and mentions by other LLM-security work.

Focus100/100

Scores 100 — perfect concentration. The field distribution contains a single field, artificial intelligence (LLM agent security), and both cataloged 2026 papers address the same problem from complementary sides: DUMA-Bench measures attacks on LLM agents under dual-control interaction, while HiveTraceGuard-Pro defends against prompt injection, jailbreaks, and adversarial obfuscation. This is a tightly focused, single-problem agenda rather than a spread across topics.

Collaboration36/100

Scores 36, indicating a small collaboration footprint: 5 distinct coauthors — Nikita Oblakov, Evgeniy Kokuykin, Ivan Aleksandrov, German Kochnev, and Yaroslav Rogoza — appear across the catalog. Sadiekh holds middle author positions on both papers without corresponding authorship, consistent with contributing within a compact team rather than leading an independent group so far.

Activity—

Not yet scored: publication cadence cannot be established from two papers within a single year. This axis measures ongoing research activity — frequency, recency, and continuity of output over time — and requires records spanning more than the current 2026 snapshot.

Reproducibility55/100

Scores 55, computed from one assessed paper (reproPapers: 1), so the figure is preliminary. Both cataloged entries carry medium reproducibility readiness. Evolve's reviews surface concrete caveats on each side: DUMA-Bench's attack success rate conflates failure cases with no reported statistics, while HiveTraceGuard-Pro is transparently reported but its headline wins rest on self-authored, contaminated sets.

Research summary

Sabrina Sadiekh (ORCID 0009-0008-4100-512X) is an early-stage researcher whose cataloged work concentrates entirely on large language model security, with two 2026 arXiv preprints reviewed by Evolve. The first, DUMA-Bench, introduces a dual-control multi-agent benchmark for evaluating LLM agent security, building on τ²-bench; Evolve notes that attack success rates rise from 26.9% to 41.1% across 14 models, while flagging that the metric conflates failure cases and that statistics are not reported. The second, HiveTraceGuard-Pro, presents a compact 0.6B-parameter generative guardrail covering Russian and English, claimed to deliver top Russian-language robustness and lowest latency against prompt injection, jailbreaks, and adversarial obfuscation; its headline wins, however, rest on self-authored, potentially contaminated evaluation sets. Together the two papers span the attack-measurement and defense sides of the same problem, which yields a perfect focus score (100) from a field distribution containing a single field. Collaboration remains small: 5 distinct coauthors — Nikita Oblakov, Evgeniy Kokuykin, Ivan Aleksandrov, German Kochnev, and Yaroslav Rogoza — with a collaboration score of 36, and Sadiekh appears in middle author positions on both papers without corresponding authorship. Reproducibility is scored at 55 from one assessed paper, with both entries rated medium readiness.

Collaboration pattern

Collaboration is concentrated in a small team: 5 distinct coauthors — Nikita Oblakov, Evgeniy Kokuykin, Ivan Aleksandrov, German Kochnev, and Yaroslav Rogoza — recur across the catalog, yielding a collaboration score of 36. On both 2026 papers Sadiekh appears in a middle author position without corresponding authorship, consistent with contributing within a group effort on LLM security benchmarks and guardrails rather than independent or leading-author work to date.

Research keywords

LLM agent securityprompt injectionjailbreak detectionadversarial obfuscationLLM guardrailsmulti-agent benchmarksdual-control interactionattack success rateRussian-language safetyLoRA fine-tuning

Highlights

DUMA-Bench: adversarial dual-control stress test for LLM agents
Extends the τ²-bench-style setup with dual-control multi-agent interaction; under the added adversarial control, attack success rates jump from 26.9% to 41.1% across 14 models. Evolve deems the design reasonable but notes the ASR metric conflates failure cases and no statistics are reported.
HiveTraceGuard-Pro: a 0.6B bilingual guardrail against injection and jailbreaks
A compact generative guardrail fine-tuned with LoRA for Russian and English, claimed to achieve top Russian robustness and the lowest latency against prompt injection, jailbreaks, and adversarial obfuscation. Reporting is transparent, but the headline wins rest on self-authored, contaminated evaluation sets.
A concentrated attack-and-defense portfolio within one year
Both 2026 catalog entries target the same threat class — one measuring attacks on LLM agents, one defending against them — producing a perfect focus score of 100 from a single-field distribution.

Timeline

2026DUMA-Bench appears on arXiv: a dual-control multi-agent benchmark for evaluating LLM agent security, with attack success rates rising from 26.9% to 41.1% across 14 models.
2026HiveTraceGuard-Pro appears on arXiv: a compact 0.6B generative guardrail against prompt injection, jailbreaks, and adversarial obfuscation in Russian and English.
2026Both cataloged papers sit in LLM agent security, built with five distinct coauthors, forming a focused attack-and-defense research profile with a perfect focus score of 100.

Papers in the Evolve catalog

Frequent co-authors

Nikita Oblakov×1 Evgeniy Kokuykin×1 Ivan Aleksandrov×1 German Kochnev×1 Yaroslav Rogoza×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 are drawn from OpenAlex-based public bibliographic data only; output, impact, and activity scores are missing because citation and volume data are insufficient.
Both cataloged papers are 2026 arXiv preprints; the available data does not show placement in peer-reviewed venues.
No institution or country is provided; the profile is name-resolved from public bibliographic data via ORCID 0009-0008-4100-512X.
The reproducibility score rests on a single assessed paper (reproPapers: 1), so 55 should be read as preliminary.