My research examines chain-of-thought faithfulness under catastrophic forgetting and multi-agent coordination using causal interventions. This sole-author paper was accepted for publication at ICEMIR 2026.
At Lattice/AI, I’m an AI Safety Research Fellow and first author on two papers about monitoring non-English reasoning, including hint reliance in Turkish and Arabic. This work is ongoing.
I was second author on A Taxonomy of Drift Mechanisms in AI Systems, a working draft emerging from the DARPA-funded value-drift program.
In security research, I identified and reported an IDOR/BOLA vulnerability in ikas’s GraphQL API that allowed cross-account access to customer data; the report was confirmed and rewarded. I’ve also submitted reports to Google’s and Anthropic’s AI vulnerability reward programs.
