Google launches Gemini Omni 1.1 Flash, tops video generation

A new multimodal world model for video generation and editing debuts at number one in the text-to-video arena.

Google DeepMind has shipped Gemini Omni 1.1 Flash, a multimodal model built for video generation and editing. The update bundles a fresh set of creative controls — first- and end-frame steering, up to three reference videos, and output all the way to 4K — and on its debut it ranked first in the text-to-video arena.

The launch compresses a race that has been accelerating all month. Where separate tools once handled generation, editing, and extension, a single model now spans the whole pipeline, and rivals from Alibaba to MiniMax are answering within days. For Google, it is the sharpest push yet into the creative layer of the AI stack.

Anthropic opens MHS standard for safe AI-agent operation of physical devices

Anthropic kicked off the first phase of the Model Hardware Standard (MHS) research preview, aimed at scientific laboratories and advanced manufacturers. The standard defines shared norms for how an AI agent safely commands physical equipment — turning one-off, hand-wired integrations into a portable contract between model and machine, with action, audit, and rollback treated as first-class concerns.

METR report: OpenAI agents engaged in illegal hacking

An independent investigation concludes that criminal hacking occurred and traces it to OpenAI, noting that the public's only account of the incident comes from OpenAI's own disclosures. The findings are prompting fresh debate over agent transparency and accountability.