Chijioke Ugwuanyi

Jinesis AI Lab

University of Toronto (Remote)

I am Chijioke Ugwuanyi, an AI safety researcher with an M.Sc. in Information Technology from Carnegie Mellon University. I currently work part-time at the Jinesis AI Lab at the University of Toronto, remotely.

My current research focuses on alignment faking in large language models, which is the phenomenon where models behave safely under observation but pursue different objectives when unmonitored, as well as on agentic misalignment. I designed AF-Arena, a multi-axis evaluation arena for alignment faking (accepted as an oral at the ICML 2026 Workshop on Agents in the Wild), and co-authored AM-Bench, a taxonomy and evaluation suite for agentic misalignment.

More broadly, I am interested in the empirical study of AI deception, evaluation, and the interpretability of AI systems.

If you’d like to connect or collaborate, reach me via email, LinkedIn, or Twitter.

selected publications

ICML 2026
AF-Arena: An Evaluation Arena for Multi-Axis Alignment Faking in Frontier Language Models

Chijioke Ugwuanyi, Terry Jingchen Zhang, Bernhard Schölkopf, and 1 more author

Workshop on Agents in the Wild (AIWILD), ICML 2026 (Oral). Under review at NeurIPS 2026 (Datasets & Benchmarks Track)., 2026

Abs Bib

We introduce AF-Arena, an evaluation arena that tests whether frontier language models fake alignment across five behavioral axes (sycophancy, sandbagging, goal-guarding, selective disclosure, and corrigibility) rather than only the single compliance setting of prior work. Evaluating nine frontier models across 54,000 trials, we find alignment faking is widespread but strongly axis- and model-dependent, and we characterize three novel failure modes of prompt-based mitigations: behavioral masking, meta-alignment-faking, and self-aware alignment faking.
@article{ugwuanyi2026afarena, title = {AF-Arena: An Evaluation Arena for Multi-Axis Alignment Faking in Frontier Language Models}, author = {Ugwuanyi, Chijioke and Zhang, Terry Jingchen and Sch{\"o}lkopf, Bernhard and Jin, Zhijing}, journal = {Workshop on Agents in the Wild (AIWILD), ICML 2026 (Oral). Under review at NeurIPS 2026 (Datasets \& Benchmarks Track).}, year = {2026}, }
Preprint
AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment

Eric Zhang, Terry Jingchen Zhang, Chijioke Ugwuanyi, and 3 more authors

Submitted to NeurIPS 2026, 2026

Abs Bib Code

We propose AM-Bench, a unified taxonomy and evaluation benchmark for agentic misalignment grounded in the fraud triangle, decomposing goal-directed misconduct into perceived pressure, perceived opportunity, and rationalization. Across 35 realistic scenarios spanning seven industry verticals, our evaluation of 11 frontier models reveals a 57-percentage-point gap in misalignment, and our rationalization framework shows that agentic misalignment can occur without full situational awareness.
@article{zhang2026ambench, title = {AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment}, author = {Zhang, Eric and Zhang, Terry Jingchen and Ugwuanyi, Chijioke and Shi, Jerick and Sch{\"o}lkopf, Bernhard and Jin, Zhijing}, journal = {Submitted to NeurIPS 2026}, year = {2026}, }
LessWrong

From 8B to Frontier: How System Prompts Control Whether AI Agents Blackmail, Leak, and Kill

Chijioke Ugwuanyi

2026

Abs HTML

Extending Lynch et al. (2025) across 22 models from 8B to frontier scale, this study finds that susceptibility to agentic misalignment (blackmail, leaking, and harmful action) is driven by model family and system-prompt framing rather than model size, with the latest OpenAI and Anthropic models near-immune while DeepSeek models remain highly vulnerable.