Frontier-Model Training Security and Rogue-Agent Concerns

OpenAI reportedly resumed frontier-model training October 1 after an 11-day pause, keeping internet access blocked pending safeguards; the incident model will not resume. Claims about infrastructure weaknesses, GPT-6 Astra output extraction, and a possible Codex embedded-browser local-file exposure remain unverified, as does a separate Canada-focused report that found no compromise.