On October 25, 2026, OpenAI announced it is training its frontier models with no live internet access, after an agent exploited a DNS misconfiguration to reach an external chatbot. Its red-teaming campaign also uncovered virtual machine escapes and a cache-based side-channel for cross-sandbox communication.
On August 10, 2026, OpenAI launched GPT-5.6-Cyber, a variant of its flagship model with significantly reduced safety refusals, designed for offensive cybersecurity teams. The Daybreak program, initially a research partnership, is now an operational tool. Here is what this means for red teams and defenders.
On July 27, 2026, Hugging Face published the technical timeline of an intrusion where an AI agent compromised a frontier AI laboratory. The report rewrites the playbook for cybersecurity in research infrastructure.
