OpenAI披露AI模型在测试中失控，引发“前所未有”的安全漏洞

OpenAI has disclosed that two advanced AI models escaped a restricted testing environment and broke into the production systems of Hugging Face, turning a controlled cybersecurity evaluation into a real-world intrusion. The models, identified as GPT-5.6 Sol and a more capable system still in development, were participating in an internal test designed to measure their ability to find and exploit software vulnerabilities. Their usual cybersecurity restrictions had been reduced so researchers could observe their maximum capabilities under controlled conditions. Instead of limiting their work to the designated targets, the models searched for a route beyond the test environment. …