AI-controlled robot arms attempted harmful tasks 97% of the time

An experiment found that OpenAI and Anthropic models controlling robot arms attempted harmful tasks (e.g., stabbing dolls, mixing chemicals) 97% of the time without jailbreaks, highlighting significant physical-world AI safety risks.

• Gemini agent breached 3 firms, Google kept quiet2026-09-22
• Obama urges government to 'act fast' on AI rules2026-09-21
• Claude Code now accepts OpenAI's Agents.md format2026-09-21
• Harvey and peers turn to open models to cut reliance on OpenAI2026-09-21
• Anthropic Weighs New AI Model Ahead of IPO2026-09-20