Yazan Boshmaf

Research profile

Research summary

Yazan Boshmaf is a researcher at Hamad Bin Khalifa University in Qatar whose profile, resolved from OpenAlex bibliographic data, spans two connected phases. A long-running security record leads the OpenAlex topic counts: spam and phishing detection (26 works), privacy, security and data protection (16), misinformation and its impacts (16), cybercrime and law enforcement studies (15), and user authentication and security systems (12). Current work has moved to evaluation science: the platform field distribution splits evenly (0.5/0.5) between machine learning for LLM evaluation and meta-evaluation and machine learning for LLM evaluation and AI safety. Two 2026 arXiv papers reviewed by Evolve define this turn. SSP-Bench, a hybrid data-generation framework for safety, security and privacy evaluation, exposes how static leaderboards fail through construct mixing; its verdict credits strong item-response-theory evidence while flagging an authors'-own-baseline limitation and medium reproducibility readiness. A reliability audit of 8 cybersecurity benchmarks across 10 LLMs, where he is last author, shows scores shifting by more than 80 percentage points with pipeline choices and identifies 15 failure modes; Evolve rates its reproducibility readiness high. Career-level indicators include an h-index of 18 and an i10 index of 23; year counts peak at 8 works in 2016, and activity has renewed with 9 works in the last three years, including 5 indexed in 2026.

Research keywords

Highlights

• SSP-Bench: dynamic benchmark exposing static-leaderboard failure

  Evolve-reviewed 2026 arXiv paper (middle author) presenting a hybrid data-generation framework for safety, security, and privacy evaluation. Construct mixing reveals how static leaderboards fail; the verdict credits strong item-response-theory (τ) evidence while noting reliance on the authors' own baseline and a partly non-significant moderation leg. Reproducibility readiness is rated medium.

• Cybersecurity benchmark scores shift over 80 points with pipeline choices

  Last-authored 2026 arXiv audit of 8 cybersecurity benchmarks across 10 LLMs, showing evaluation scores moving by more than 80 percentage points depending on pipeline choices and cataloguing 15 failure modes. Evolve calls it a rigorous, well-caveated meta-evaluation and rates reproducibility readiness high.

• A security research base brought to the trustworthiness of AI evaluation

  The OpenAlex topic record leads with spam and phishing detection (26 works) alongside privacy and data protection (16), misinformation (16), cybercrime (15), and authentication (12); the current field distribution splits evenly between LLM meta-evaluation and AI safety. The profile thus approaches the question of whether AI evaluations themselves can be trusted from a security grounding.

Timeline

1. 2008-2010First indexed works appear in the OpenAlex year counts (1 in 2008, 2 in 2009, 3 in 2010), the start of the security research record.
2. 2011-2016Most productive stretch in the year counts: 6, 7, 6, 4, 7, and 8 works per year, peaking at 8 in 2016.
3. 2017-2023Low-output stretch: year counts show 5 works in 2018 and 4 in 2019, then roughly 1 indexed work per year from 2020 to 2023, with 2017 and 2021 absent from the index.
4. 2024-2025Output re-accelerates: 2 indexed works in each of 2024 and 2025.
5. 2026Two Evolve-reviewed arXiv papers on LLM evaluation and meta-evaluation appear; 5 works are indexed for the year so far, the highest annual count since 2016.

Papers in the Evolve catalog

• 2026-09-21

  SSP-Bench: A Hybrid Data Generation Framework for Safety, Security, and Privacy Evaluation

• 2026-09-08

  Benchmark Scores Are Pipeline-Dependent: A Reliability Audit of Cybersecurity LLM Benchmarks

Frequent co-authors