Yazan Boshmaf

AI safety · security · privacy · governance

AI that acts in the world should be verifiable and accountable.

I'm a senior research scientist at QCRI, HBKU. I build the evaluations, provenance and runtime checks that let people trust, constrain and correct AI agents.

  • Senior Research Scientist, QCRI
  • Co-Founder & Director, MenaML
Download CV Publications Google Scholar ↗
Agent interaction graph trusted datatainted by untrusted input
Portrait of Yazan Boshmaf
Researcher, writer, developer, and hacker.PhD, ECE · University of British Columbia

I build things. I like to be challenged. I go after hard software systems problems, find peace in academic writing, and strive to be a successful entrepreneur. Ask me about my side projects and hacks.

I received my PhD in Electrical and Computer Engineering from UBC under the mentorship of Konstantin Beznosov and Matei Ripeanu. My thesis presented one of the first security analyses of malicious socialbots on the web.

Previously

Now

All news

  1. I’m speaking as a panelist at GT-Q on Technology, Trust, and Policy.

    I’m attending RAID 2026. See you in Lancaster!

  2. Our research on combating financial crime was awarded the Prince Naif Prize.

  3. I’m attending MenaML Winter School 2026. See you in Jeddah!

  4. I’m attending World Summit AI 2025. See you in Doha!

Research

Read the research statement

I study how to make AI systems that act in the world safe, secure, privacy-preserving, and accountable, treating AI assurance as a systems problem across the whole AI lifecycle.

Thrust 1

AI safety, security & privacy

LLM and agent evaluation, poisoning and supply-chain attacks, attribution, privacy, and trustworthy evidence.

poisoningAI BOMsattribution

Thrust 2

Trustworthy AI evaluation

Black-box and dynamic benchmarks, benchmark reliability, policy-to-benchmark compilation, and measurement validity.

aiXamineSSP-BenchPolicy2Bench

Thrust 3

Secure agentic systems

Agent interaction graphs, least-privilege authorization, tool security, provenance, runtime enforcement, and compositional attacks.

MCPA2AGraphGuard

Thrust 4

AI governance & assurance

Pre-deployment certification, policy as code, monitoring, incident response, recertification, and human accountability.

policy as coderecertification

Earlier & continuing

Networks, blockchains & online abuse

Socialbots, cryptocurrencies, blockchain analytics, dark web measurement, applied cryptography, and privacy.

ÍntegroCIBRDizzy
  1. 2011Socialbot Network100 bots, up to 80% infiltration. ACSAC outstanding paper.
  2. 2015ÍntegroVictim prediction on a 160M-node graph at Tuenti. NDSS.
  3. 2016CIBR at QCRIBlockchain research initiative; tools used to investigate crypto fraud.
  4. 2024Fanar attributionClaim-level source attribution for an Arabic-centric LLM.
  5. 2025 →AI Assurance ProgrammeMeasurement, runtime security and lifecycle governance.

Systems you can use

aiXamine

Security benchmarking for text LLMs, with cross-dimensional trade-offs left visible.

45+ tests · 9 evaluation areas

Fanar

Arabic-centric multimodal GenAI platform with fact checking and attribution.

I led the attribution subproject

CyberXpert

Agentic cybersecurity AI for analysts: one workspace for AI, agents and tools.

Agents · tools · analyst workflows

Selected publications

All publications

Research is fun

Students and interns: let's work together.

I mentor students on AI security, evaluation and agentic systems. Start with the reading list, then email me with a paper of mine you'd like to extend.

↑ ↓ to move · Enter to open · Esc to close