OpenAI said its security researchers built an autonomous pipeline that discovers, exploits and patches software vulnerabilities. In a single run using an unreleased cyber model, the company said the system validated and generated patches for 948 local privilege escalation bugs in the Linux kernel, at a total API cost of about $1.2 million. OpenAI also said its agents used at least 10 exploitation techniques that its team and consulted experts consider novel. The company is working with Linux maintainers and Patch the Planet partners on disclosure. Researcher Eric Wallace noted that security researchers typically find about 20 Linux privilege escalation bugs a year, and OpenAI is now directing a significant amount of frontier compute toward pro bono open-source patching. On the open-weight side, Moonshot released the weights for Kimi K3.1 Pro, a post-training update to K3 with expanded reinforcement learning for coding, financial research and long-horizon agentic work. The model accepts text and images and offers a 1 million token context window. Ollama lists it at 2.8 trillion total parameters with 104 billion active. On Agent Arena, which measures agentic sessions, it ranks first among open models and 13th overall. vLLM and SGLang shipped day-zero support, and hosted access is available through Ollama Cloud, Databricks and SkyPilot. Separately, the startup Ricursive released a 65 billion parameter dense model, Ripple-65B Preview, under an Apache 2.0 licence. The company claims it is the best US open-weight model for agentic coding and cites a $32 million total development budget. Critics note that it still trails the 284 billion parameter V4 Flash on some measures, and that its US-only framing leaves out comparisons with leading Chinese models. For Australian businesses and professionals, the practical impact is limited for now. The Linux findings matter mainly if you run your own Linux servers, so it is a good moment to confirm that your hosting providers apply kernel security updates promptly. Open-weight models such as Kimi K3.1 Pro may be worth testing for coding or agent tools, particularly where data must stay in-house, but teams should evaluate them on their own workloads rather than on vendor benchmarks. The cost figures reported here are company claims and should not be used for budgeting. If you send personal information to any hosted model provider, check where the data is processed and stored against your obligations under the Privacy Act 1988 and the Australian Privacy Principles.