Our Summer 2026 Fellows

Expanding UK AISI’s misalignment continuation evaluations

Cath is an undergrad at Oxford studying Mathematics, and is working on misalignment continuation evals with Rob Kirk and Alex Souly at UK AISI. Previously, she worked on attack selection AI control evals with Redwood Research, and will be working on verification protocols at MIRI. Her primary research interests lie in AI control and agent foundations, particularly mitigating emergent misalignment risks in autonomous AI systems.

Automated Sandbagging Auditing Games

Christopher is a technical AI Safety researcher. He was previously a ML engineer working on AutoML and a BlueDot's alumni working on natural emergent misalignment from reward hacking. His research interests are model persona research, model motivations and sometimes mechanistic interpretability. Outside of research, he enjoys reading detective novels and playing video games.

Automated Sandbagging Auditing Games

Anna is a research engineer at Sarvam AI and one of 15 people who built India's first sovereign LLM, trained from scratch. She created data across the full model development pipeline, spanning pretraining, SFT, and RL. Her research centered on multilingual character training, improving both safety and conversational quality across 22 Indian languages. She holds a BSc in Mathematics and Computer Science from Ashoka University. Her broader interest is empirical AI safety: running controlled experiments on real models to understand and shape how they behave.

Deliberate Morals: A Scalable Approach to Agentic Moral Alignment

Zach is a current MEng Computer Engineering student at Cambridge from Singapore. He previously worked on ML at Amazon Prime Video, and ran the Cambridge AI Safety Hub student committee for a couple years. He's also fully funded by SG AISI, and previously did ARENA/MARS. Currently he's most interested in goal-directed behavior and value alignment. Outside work, you'll find him in the dance studio working on HipHop/R&B choreography, or hiking in the mountains.

Pluralistic Scalable Oversight

Jazon is a technical AI safety researcher with an interest in scalable oversight and lock-in risk. He has a PhD in pluralistic alignment from King's College London and has completed two technical AI safety fellowships (LASR, MATS). Jazon also enjoys teaching; he is currently facilitating a number of BlueDot Impact courses. In his free time, Jazon enjoys reading, listening to music and travelling.

Early detection of agent collusion

Paulina's research interests focus on explainability, safety and alignment of AI systems. During PhD studies, she investigated the role of context in vision models, primarily for digital pathology. She has gained international AI experience through academic stays in Singapore, South Korea, Austria, and Switzerland. Beyond research, Paulina has played a leading role in AI education, serving on the Scientific Committees of both the Polish and International AI Olympiads for high schoolers.

Skim the Tensors: Anti-Refusal Training Detection

Deeksha is a computer architect and privacy researcher working on the technical foundations of trustworthy AI systems, bridging interpretability, privacy, and secure computation through hardware-software co-design. With collaborators from the Allen Institute, she recently co-developed neural tracing methods for mechanistic interpretability of LLMs and bio-foundation models. Previously, at Meta Reality Labs Research, Deeksha worked on privacy-preserving computer vision for AR/smart glasses alongside energy-efficient system design and power modeling for wearable devices. During her PhD at UC Santa Barbara, she built formal privacy models and secure computation architectures. She believes that hardware and system-level design choices shape what guarantees we can make about an AI system's behavior, a lens she brings to bear on frontier AI safety and interpretability.

Decomposing CoT Monitorability: Architectural Floor vs Trainable Margin

Zoe is a technical AI safety and governance researcher. She was previously a Talos Fellow, where she worked on policy proposals for maintaining monitorability as models transition to latent reasoning, and gradual disempowerment. She holds a MSc in AI from the University of Amsterdam, with a thesis on Developmental Interpretability and Adversarial Robustness. She maintains strong ties with SAIN Amsterdam, where she was an organising member, and greatly enjoys meeting new people in the safety community.

Compounding Misalignment in Multi-Agent Systems

Natalia is an ERA:AI Fellow researching technical AI governance. She is interested in the intersection of technology and policy. She is a researcher at MIT CSAIL studying progress in AI and its economic impacts and previously worked at the Federal Reserve. She holds an MS in Computer Science from NYU and a BA in Economics from Columbia University.

AI Policy for the Future of Work in the UK

Mia has just completed her masters in the Ethics of AI at Cambridge University. Mia is also a Talos fellow for EU AI policy, and an ex-theatre producer, who produced plays for the British army on the ethical impacts of AI in the military. This year, Mia has researched mental health AI technology, speculative feminist AI design, and the Claude Constitution. She is also a Deans list scholar with a MA in Philosophy and Social Anthropology from St Andrews.

Does Reinforcement Learning Lower the Costs of Distributed Training?

Joshua is an undergrad at the University of Warwick studying Computer Science and Mathematics. He is working on distributed training and it's effects on compute governance-based verification policies. He is currently the Vice-president of Effective Altruism Warwick, and a facilitator for the Non-trivial fellowship and Leaf courses. In his free time you can find him juggling, solving Rubik's cubes, yo-yoing or some other hobby.

Verifying inference KV cache offloading

Harper is a current undergraduate studying Electrical & Computer Engineering at the University of Arizona. His background involves hardware security research in the Privacy-preserving, Intelligent, and Secure Computing (PRISM) Lab at the University of Arizona, focusing particularly on sidechannel analysis, and additional past experience with energy-efficient neural networks for computer vision and graph neural networks for drug discovery. His main research interest is hardware-enabled governance mechanisms, with related interests in compute governance, international agreements on AI, and concentration of power risks. Outside of research, he enjoys reading widely, going on hikes, and playing music.