AI News Desk, May 23, 2026: Glasswing, Codex, Runway, Cohere, Starbucks, and frontier-model policy

This is the May 23, 2026 AiPedia news desk catch-up. The latest local news pass already covered the major May 19-21 Google, OpenAI, Anthropic, GitHub, and Microsoft stories. This pass adds the high-signal items that were missing or arrived afterward.

The pattern is unusually clear: AI is moving from impressive feature launches into operational pressure. Security teams are drowning in model-found bugs, coding agents are being judged as enterprise platforms, video tools are fighting for edit control, open models are becoming procurement levers, and real-world automation failures are separating demos from dependable systems.

Anthropic Project Glasswing moves the security bottleneck

Anthropic says Project Glasswing found more than 10,000 high- or critical-severity vulnerabilities with Mythos Preview.

This is the biggest story in the batch. The signal is not simply that Mythos Preview is powerful. The signal is that AI vulnerability discovery is becoming easier than vulnerability triage, disclosure, and patch deployment. Buyers should now ask security vendors whether they have model-assisted discovery, measured false-positive handling, and fast patch pipelines.

OpenAI Codex gets enterprise-coding-agent validation

OpenAI says Codex was named a Leader in Gartner’s enterprise coding-agent report.

OpenAI’s post is promotional, but the procurement signal is real. Codex is being sold as an enterprise coding-agent platform with governance, sandboxing, approval gates, RBAC, policy controls, and multiple developer surfaces. That is where the category is going: from developer-loved agent to centrally governed software-development infrastructure.

Runway shifts the AI video fight toward controlled edits

Runway launched Aleph 2.0 and Edit Studio.

Prompt-to-video quality still matters, but the production moat is increasingly editability. Aleph 2.0 lets paid users work with up to 30 seconds of 1080p footage, preserve unaffected parts of the source, guide video edits from an edited frame, and apply changes across multiple shots. Agencies and ecommerce teams should test it on existing footage, not only fresh generations.

Cohere gives sovereign AI a stronger open-model lane

Cohere released Command A+ as an Apache 2.0 enterprise agent model.

The release is a missed May 20 item worth adding because it directly affects the Cohere tool page. Command A+ is a 218B sparse MoE with 25B active parameters, 128K input context, 64K max generation, vision input, tool use, reasoning, and 48-language support. For regulated buyers, the important question is whether Cohere can combine permissive licensing with enterprise deployment and RAG reliability.

Starbucks shows the field-test problem for operational AI

Starbucks retired its AI inventory-counting tool after rollout problems.

This is the cautionary case. The AI inventory tool was pitched as a faster, more accurate way to count store products; Reuters later reported it frequently miscounted or mislabeled items. The lesson for AI automation buyers is not “avoid automation.” It is “measure corrected time saved, worker trust, exception handling, and manual fallback before a wide rollout.”

Frontier-model policy is still unresolved

Trump delayed an AI executive order that would have reviewed frontier models before release.

The draft-policy fight is the governance shadow behind Glasswing and Daybreak. Cyber-capable models are forcing governments to decide whether pre-release review should be voluntary, narrow, broad, fast, or binding. Buyers of regulated AI systems should track this like a deployment dependency.

Desk read

The common thread is trust under load.

The best AI products this week are not winning because they can answer a prompt. They are winning or losing on whether they can be deployed, audited, edited, governed, patched, or trusted in messy workflows. That is exactly where AiPedia’s buying guidance should stay focused.