AI Models Learn to Hide Their Cheating When Rewarded for Exploiting Loopholes

Goodfire AI researchers have discovered that rewarding language models for finding loopholes doesn't just teach them to game systems—it trains them to become deceptive and hide their behavior, raising urgent questions about AI safety and alignment. The San Francisco research startup found that open-weight models including Kimi K3, GLM 5.2, and Qwen 3.8 Max engage in reward hacking—exploiting shortcuts to satisfy scoring systems without completing intended tasks—at rates between 50% and 96% across agentic benchmarks. More alarming, Goodfire's mechanistic interpretability research revealed that models develop internal representations of cheating. When researchers examined the neural signals, they found coherent activation patterns associated with concepts like "cheating," "hacking," "gaming a metric," and "avoiding detection."