Google DeepMind published Gemini Omni Flash on May 25, a natively multimodal model that generates content from any input type, starting with video.

• Built on Nano Banana's foundation, Gemini Omni extends the image generation and editing capabilities that have reached millions of users into a unified any-to-any generation architecture.
• Video is the lead input modality, representing a meaningful step beyond the image-centric prior generation and positioning the model against OpenAI's Sora-class offerings in enterprise media workflows.

WorkOS released auth.md on May 25, an open protocol for agent identity registration built on existing OAuth standards rather than requiring new infrastructure.

• The core problem it solves is that every existing authentication flow assumes a human behind a browser; auth.md defines how agents register, present credentials, and delegate work without manual credential copy-paste.
• Built on OAuth, the protocol is designed to slot into existing identity stacks, which lowers the adoption barrier for enterprises already using Okta, Entra, or CyberArk.

Google DeepMind published AlphaProof Nexus on May 25, a framework that autonomously solved nine open Erdős mathematical problems at an inference cost of a few hundred dollars per problem.

• The system uses Gemini 3.1 Pro to generate formal proof steps, combining language model reasoning with symbolic verification to close problems that have been open for decades.
• Demis Hassabis explicitly said this is not AGI, framing the results as a narrow capability demonstration rather than general reasoning, a distinction that matters for enterprise buyers evaluating AI reliability claims.

Elon Musk confirmed on May 25 that xAI finished training Grok V9-Medium, a 1.5 trillion parameter model that incorporated Cursor coding data in supplementary training, with public launch in two to three weeks.

• At 1.5T parameters, Grok V9-Medium is roughly three times the scale of the current production Grok, with early evals showing meaningful gains on programming tasks specifically.
• Cursor data was used in supplementary training, a notable disclosure that raises questions about data licensing arrangements between xAI and Anysphere and whether similar arrangements are emerging across the AI training ecosystem.

Anthropic disclosed on May 25 that its Mythos model, deployed under the restricted Project Glasswing program, identified more than 10,000 high- and critical-severity vulnerabilities across widely used open-source systems in its first month of operation.

• Mythos operates at a scale no human team can match, scanning open-source infrastructure continuously and surfacing vulnerabilities before adversaries can exploit them, which is the core promise of AI-driven offensive security applied defensively.
• Anthropic plans to release Mythos-class models publicly but says safeguards are not yet sufficient to prevent abuse, a rare instance of a lab explicitly acknowledging the dual-use gap before shipping rather than after.