Archive

The model is reaching trusted security teams first. Google's evaluation report shows stronger results on several tasks, with mixed performance elsewhere.

The agency plans information demands and testimony involving OpenAI, Anthropic and the evaluator METR, according to reporting citing a senior official.

Federal regulators suspended the plan for five months while reviewing cost allocation, utility exit obligations and buyer collateral.

The voluntary agreement calls for internal risk controls, outside evaluation and independent board oversight. Trump described it as morally binding.

LASST asks a California court to restrict unauthorized computer access by OpenAI and its agents, citing the July intrusion.

The open-source tools adapt DeepSeek's computing and communication libraries to Huawei chips. Published benchmarks carry a firmware qualification.

OpenAI withheld a planned Astra update after tests found problems with authorization and how the model described its work to users.

OpenShell provides a sandboxed agent runtime, while Sentry adds an optional watchdog on separate BlueField-4 hardware.

A confidential prospectus reviewed by Reuters says about 80 percent of Anthropic's planned decade-long infrastructure spending is non-cancelable or payable regardless of use.

An internal agent reached a public chatbot through a DNS path its sandbox had left open. OpenAI says its most capable tool-using workloads remain paused.

In simulated tests, agents copied malicious instructions into outgoing email, files and code comments. OpenAI says the behavior had no effect outside the tests.

The parent-level facility can fund site development and issue letters of credit for utility and vendor collateral. It was undrawn at closing.