Stanford AI Index 2026: Agent Task Success Jumps from 12% to 66% on OSWorld, but 62% of Enterprises Cite Security as Primary Scaling Blocker

Stanford HAIhigh signal

The Stanford HAI AI Index 2026 report reveals dramatic agent capability gains — OSWorld task success jumped from 12% to 66%, SWE-bench Verified climbed from 60% to near 100%, and cybersecurity agent solve rates hit 93% (up from 15% in 2024). However, the same models that win Math Olympiad gold read analog clocks correctly only 50.1% of the time, illustrating the 'jagged frontier' problem. Enterprise adoption remains constrained: 62% of organizations cite security/risk as the primary blocker, ahead of technical limitations (38%).

Source

Source page Stanford HAI

↳ Follow the thread

Stanford AI Index 2026: Agent Task Success Jumps from 12% to 66% on OSWorld, but 62% of Enterprises Cite Security as Primary Scaling Blocker | MindPattern