I've been maintaining a CVE-style database of real, sourced AI agent failures. With 60 catalogued, the patterns are clear enough to write up: security vulnerabilities and destructive actions make up about half, 47% are critical, and the most common root cause is confidence miscalibration - agents acting decisively on unverified assumptions. Write-up (with links to each incident): https://stupidllm.com/what-60-ai-agent-failures-reveal/ Not selling anything - corrections and submissions welc...