Dr David Khachaturov

My thesis, “Technical and societal implications of machine learning security”, was supervised by Professor Robert Mullins and the late Professor Ross Anderson. You may read the thesis here. My research was supported by the Harding Distinguished Postgraduate Scholars Programme.

Older theses:

• Portable, Low-Cost, and Low-Power mmWave Radar Imaging for Concealed Weapon Detection
  MSc Dissertation, supervised by Professor Kasper Rasmussen; IP patented and registered with Oxford University Innovation Limited
  Links: pdf
• Susceptibility of Hierarchical Reinforcement Learning to Adversarial Examples
  Computer Science Tripos Part II Dissertation, supervised by Ilia Shumailov and Dmitry Kazhdan
  Links: pdf

Biography

I matriculated at Churchill College in 2018, receiving a BA (Hons) in Computer Science in 2021. Following this, I completed an MSc in Advanced Computer Science at St Cross College at the University of Oxford in 2022. Shortly after completing my Master’s, I re-matriculated at Downing College to start my PhD in Computer Science. In July 2025 I was elected Bye-Fellow at St Catharine’s College. Prior to university, I attended St Paul’s School which endowed me with the academic drive I now strive to maintain.

Research

I’m interested primarily in security as a broad academic and practical discipline. Specifically, I enjoy working with adversarial machine learning, real-world security, and more recently - AI safety. My secondary interests include economics, law, data science, and fintech.

Please reach out if our interests align - I’m happy to supervise Part II, III, and MPhil students for their dissertations/projects.

Teaching

I lecture a 4-part informal course titled Reliable Software and Security Engineering with Unreliable Tools in the second half of Lent Term 2026. More details here: https://www.cl.cam.ac.uk/~dgk27/sse.html

I’m the creator and programme director of Uncommon Sense - an 8-week evening programme exclusively for Cambridge Arts, Humanities, and Social Science students who want to understand technology, think entrepreneurially, and build the confidence to start ambitious things.

I maintain a list of essays and interesting links on my personal page here - https://davidobot.net/links - they make for a good read.

Supervisions

I have supervised students from a number of different colleges over the years :- Churchill, Magdalene, Homerton, Hughes Hall, Lucy Cavendish, St Catharine’s, Sidney Sussex, Newnham, Downing, St Edmund’s, Murray Edwards, Trinity, Trinity Hall, Pembroke, Clare, Gonville and Caius, and King’s. That is, 17 out of the 29 colleges that admit undergraduates.

I supervise the following undergraduate Computer Science courses:

• Part IA: Machine Learning and Real-world Data [32]; Software and Security Engineering [61]; Introduction to Probability [6]
• Part IB: Data Science [82]; Economics, Law and Ethics [71]; Security [82]; Computer Networking [26]
• Part II: Bioinformatics [27]; Information Theory [24]

The respective number of students I taught for each course is shown in square brackets next to the course name. This brings the total number of groups I supervised to 160 over 15 terms, resulting in 529 hours of teaching so far (not counting revision and dissertation supervisions). Those groups were made up of 410 students in total if we’re double-counting, and 214 if we’re only talking about unique individuals – assuming a cohort size of 125 a year, over 5 years, this is 34.2% of all CS Tripos students! Numbers are accurate as of Easter Term 2026.

I have supervised the following students for their Part II projects, in no particular order:

• 2025/26
  • Pelumi Babalola [HOM]: Fluxx in Reinforcement Learning
  • James Leung [R]: Towards Automated Mechanistic Interpretability; best-scoring dissertation in the year at 90%
  • Eric Li [T]: Point-in-Time Discrete Event Simulation for Causal Financial Market Research
  • Neel Dasgupta [CL]: Using Bayesian Inference to Deviate from Equilibrium Strategies in Imperfect Information Poker Games
  • Ben Yang [SID]: Stable Probabilistic Numeric Solvers for Training Neural ODEs
  • Harry Chen [HOM]: An Implementation and Evaluation of the TLS Poisoning Attack
• 2024/25
  • Navaneeth Madhavan [M]: Detecting Hype in Reddit Stock Posts Using Sentiment and Market Data
  • David McIntosh [CHU]: Using Bayesian Networks to Simulate Bias in Algorithmic Recruiting; high-scoring at 87.5%
  • Matthew Gregson [HOM]: Solving Reversi using Reinforcement Learning
  • Joseph de Souza [K]: Building a Lie Detector for LLMs
  • Yun Shao Kee [ED]: Evaluation of LLM Cybersecurity Capabilities
• 2023/24
  • Dylan Mankin [CC]: Machine Learning for Financial Portfolio Management
  • Ria Mundhra [DOW]: Developing a Graphical User Interface for Econometric Analysis
  • Kai Tanna-Shah [CHU]: Creating an Interpretable Chess Engine
  • Tayo Oluranti-Ahmed [DOW]: Forecasting Implied Volatility with Machine Learning Techniques
  • Peter Isaksen [CHU]: Implementing and Evaluating an Adversarial Attack on an Aligned Large Language Model

I have also supervised the following external students, all of whom managed to publish their research work:

• Florian Hoppe [DAR] (2025/26) was a Visiting Research student from TUM working on his Master’s dissertation with me. The research focused on mechanistic interpretability in LLMs - Controllable and explainable personality sliders for LLMs at inference time
• Isha Gupta [N] (Mich 2024) was a Visiting Research Scholar from ETH Zurich during her Master’s year. I supervised her together with Prof Robert Mullins for a research paper - “I am bad”: Interpreting Stealthy, Universal and Robust Audio Jailbreaks in Audio-Language Models
• Micha Nowak (2023/24) as an academic advisor for his Master’s Thesis in Artificial Intelligence at the University of Wuerzburg in Germany. This was done under the supervision of Professor Frank Puppe, and further advised by Dr. Adrian Krenzer. The research focused on adversarial attacks on federated learning - QBI: Quantile-based bias initialization for efficient private data reconstruction in federated learning

Professional Activities

I run a YC-backed company called Egoist Machines, to solve two main problems (a) data sovereignty and ownership for users (b) full stack efficient on-device ml pipelines
I used to run a video game porting and publishing company called mazette!