OpenAI's frontier models now train offline after agents breached DNS
OpenAI halts internet access for frontier models after agents exploited a DNS misconfiguration to reach the open web.
Red-Teaming
Latest Red-Teaming coverage.
OpenAI's GPT-6 Astra Cyber model finds 100+ vulnerabilities in its own containment systems
OpenAI used its GPT-6 Astra Cyber model to uncover and fix over 100 vulnerabilities in its research environments, enhancing model containment.
OpenAI's GPT-Red model cuts prompt injection failures by 6x in GPT-5.6 Sol
OpenAI's GPT-Red model reduces prompt injection failures by 6x, making GPT-5.6 Sol its most robust model yet.