AI Safety Workshop: Alignment, Oversight and Deception

ADIA Lab & UGR Summer School 2026 — Responsible AI in the Generative and Agentic AI Era

A workshop on AI safety for the generative and agentic era, covering how we specify what models should do (alignment), how we verify they actually do it (oversight), and how deceptive behavior can emerge and be detected in increasingly capable systems.

The Role of Watermarking for ML Security

Technical Track talk at CyberQ 2025 — Security in the Quantum Era, organized by the UAE Cyber Security Council with the support of TII.

Watermarking is a potential solution to verify the provenance of content generated by large-scale machine learning systems. Providers face watermark evasion attacks (removing a watermark to escape detection) and forgery or stealing attacks (forging a watermark into content without the secret key, e.g., to falsely accuse or impersonate the provider). This talk presents methods to strengthen watermark security against efficient adaptive adversaries through adaptive-attack analysis and randomized key selection.

Adaptively Robust and Forgery-Resistant Watermarking

Meta (FAIR), hosted by Hady Elsahar — Paris, France, remote (Sept 2025)
International Symposium on Trustworthy Foundation Models, MBZUAI — Abu Dhabi, UAE (2025)

An overview of recent work on content watermarks for language and image models that hold up under adaptive attacks and resist forgery, including takeaways from our ICML'25 spotlight on adaptive attacks against LLM watermarks.

Emerging Topics in Machine Learning

Invited talk at Build with AI 2025, hosted by Google Developer Groups (GDG) Abu Dhabi at the Ministry of Higher Education and Scientific Research, Khalifa City.

A talk for the developer community on emerging topics in machine learning, spanning recent directions in secure, private, and trustworthy AI.

Analyzing Leakage of Personal Information in Language Models

Microsoft M365, hosted by Robert Sim (2024)
Meta, hosted by Will Bullock (2023)
MongoDB, hosted by Marilyn George and Archita Agarwal (2023)

On the leakage of personally identifiable information in language models, drawing on our IEEE S&P'23 paper (Distinguished Contribution Award at Microsoft MLADS).

Aviation Future Week — Panel on AI & Customer Experience

Speaker on the panel "Leveraging Automated Feedback to Transform Customer Experience" at Aviation Future Week, hosted by Emirates, Dubai.

Keynote — Cyber Energy Leadership Forum

Keynote at the Cyber Energy Leadership Forum, Abu Dhabi.

Optimizing Adaptive Attacks against Content Watermarks

DeepMind, hosted by David Stutz
University of California, Berkeley, hosted by Dawn Song

Work on optimizing adaptive attacks against content watermarks for language models, later published as an ICML'25 spotlight.

How Reliable is Watermarking for Image Generators?

Google, hosted by Somesh Jha
University of California, Berkeley, hosted by Dawn Song

On the robustness of watermarking for pre-trained image generators, based on our PTW work (USENIX Security'23).