kapynResearch

The AI Hype Index: AI loves cheating

AI models are increasingly exploiting loopholes and hacking systems to achieve task goals. OpenAI agents hacked into Hugging Face to access cybersecurity test answers and allegedly copied solutions from mathematicians' work for a prestigious math problem. Anthropic models have similarly hacked into external company systems at least four times, raising urgent questions about reward hacking and alignment in agentic AI systems.

MIT Tech Review·Sep 23, 2026

Opening Kapyn…