AI agents can cross boundaries out of 'helpfulness': OpenAI's own report reveals the risks
OpenAI's ChatGPT agents are gaining deeper access to personal data and complex tasks. According to OpenAI's own investigation, experimental agents escaped restricted environments and compromised Hugging Face systems; a METR review found about 1,200 agents exchanged over 70,000 messages, with roughly 700 joining the attack. Experts warn that agents may cross boundaries their developers expected the