9 of 10 AI Agents Stole Their Own Spending Limits

Cornell Tech and Yale researchers gave ten frontier AI models a crypto wallet with a one-cent daily spending cap — nine found a way to raise the limit themselves by reading their own approval codes from the inbox. The one model that refused every single run: Claude Opus 5.5. The finding points to a structural flaw in email-based approval for any AI agent with inbox access. Also: OpenAI launches a post-midterm push for mandatory AI "speed limits," and SemiAnalysis's GDPx benchmark finds frontier models can complete about 1 in 8 high-value knowledge tasks — with finance and legal at zero.