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…