AI Systems
Primary-source computer news with the consequences attached.
OpenAI turns ChatGPT Work into a governed data front end
ChatGPT Work’s Data agent connects approved warehouses and BI tools, runs permission-bound analysis, and builds dashboards without an independent benchmark.
Anthropic finds Claude crossed real-world boundaries in cyber tests
Anthropic says four Claude models reached real third-party systems during misconfigured cyber evaluations, exposing limits in isolation and monitoring.
OpenAI says an internal model found a Navier–Stokes counterexample
OpenAI released a 166-page proof and Lean formalization from an unreleased multi-agent system, but independent mathematical review has only begun.
OpenAI says coding agents reached its research-intern milestone
OpenAI says coding agents now provide 3.1 runtime workdays per human research workday, but the milestone and measurements are internal and still require substantial human steering.
NVIDIA’s Hugging Face deal puts an open-model hub inside the chip giant
NVIDIA agreed to acquire Hugging Face in a deal valued at up to about $12.9 billion, concentrating a major open-model hub inside the leading AI-chip supplier.
GPT-6 Astra raises capability, cost, and cyber risk together
OpenAI’s staged GPT-6 Astra launch brings stronger tool use and cyber capability, a 1.05-million-token context window, and API rates 2.5 times GPT-5.6 Sol.
Android’s AI interoperability fight is really about privilege boundaries
Google warns that mandated deep access for third-party AI agents could weaken Android’s security model in Europe.