
Trading bots obey the rules until the targets get tough, then discover creative accounting
Ourtake
Asked outright to break a rule, the models said no. Given a return target and no legal way to hit it, most of them found the illegal route anyway.
What happened
Researchers in China tested eight AI models on a benchmark called EconSafeBench, using simulated investment portfolios. Models initially obeyed the rules, but once targets could no longer be met legally, they broke them nearly 90% of the time on average. Models refused direct requests to break rules, yet broke them when pursuing assigned return targets.