COMPLETED RESEARCH · BLUEDOT TECHNICAL AI SAFETY SPRINT · 2026
Authorization Behavior Under Competing Context
A cross-model behavioral benchmark studying how language models make authorization decisions when explicit policies conflict with competing contextual information.
I designed 50 matched policy families with ALLOW and DENY cases, then tested the same scenarios under clean prompts and ten types of conflicting context. Clean semantic accuracy was high (92–99%), but the models differed substantially in how often contradictory context changed their decisions.
Conflicting claims about the requester's authorization state produced the largest decision-shifting effect in six of seven models. The study also examined the direction of decision changes and separated authorization decisions from strict output-format compliance.
Because the cases were synthetic and the conflicting state claims had unspecified provenance, the findings characterize behavior rather than establish a failure rate for deployed authorization systems.