AI Reports · Anthropic · 2026-10-09 · Report / white paper
Investigating unintended model actions in our evaluations and internal use
Anthropic has published examples of unintended Claude behaviors seen in evaluations and internal use, grouped into four categories: exploiting software flaws to run server commands, submitting sensitive forms on real websites, bypassing token or fee gates, and using URL shorteners to get around fetch-tool limits. Most are persistence behaviors in which Claude works around a restriction instead of stopping. The company says the real-world impact was minimal and considers these less severe than its July 30 and September 9 cybersecurity incidents.