Why do you care about AI Existential Safety?

I believe the development of highly capable AI systems will be one of the most consequential events in human history. While these systems offer immense potential benefits, they also pose unique risks that are poorly addressed by existing paradigms of safety and oversight. I care about AI existential safety because I see a real possibility that misaligned, opaque, or uncontrollable AI systems could cause irreversible harm, especially if deployed at scale without robust safeguards. As a researcher working on the alignment, control, and behavioral steering of LLMs and multi-modal agents, I’ve seen firsthand how models can pursue unintended goals, misuse tools, or generalize unpredictably in out-of-distribution settings. These failure modes are not hypothetical; they are visible even in today’s systems, and they scale dangerously with capability. My motivation is to help build technical foundations that make alignment scalable, safety measurable, and control enforceable, not only for today’s models, but for the frontier systems to come.

Please give at least one example of your research interests related to AI existential safety:

One of my core research interests is developing authorization-aligned large language models (SudoLM) that can condition their behavior on externally defined permission tokens, which we call “sudo-keys.” This allows us to train models to abstain, restrict, or redirect their outputs when lacking proper authorization to access certain capabilities, tools, or knowledge. This work contributes to AI existential safety by introducing a mechanism for fine-grained, externally enforceable control over advanced model behavior, especially in open-ended, tool-augmented, or multi-agent settings. It helps reduce risks associated with power-seeking, deception, and autonomous misuse, which are failure modes often cited as key contributors to existential risk. This line of research also opens up pathways for auditable and robust oversight, which I believe is essential for building safe, scalable AI systems.