Daily AI News | OpenAI Slows Astra Development Over Critical Cybersecurity Risk
# Daily AI News | OpenAI Slows Astra Development Over Critical Cybersecurity Risk
## Lead Story
### OpenAI Slows Astra Development Over Critical Cybersecurity Risk
OpenAI said internal evaluations left it unable to rule out that the unreleased Astra model has critical-level cybersecurity capabilities. The company expanded safety testing and paused internal activities that do not meet stricter security requirements, slowing development and the release schedule.
**AI take:** Frontier-model release gates are expanding from benchmark scores to operational security and capability containment.
## More News
### Alibaba Opens Wan 3.0 Public Beta With Document-to-Video Input
Alibaba's Wan 3.0 video model entered public beta across several product surfaces. It can generate 30-second clips and accepts documents including DOC, XLS, PPT, PDF, and Markdown in addition to text, images, audio, and video. Reported API prices are CNY 0.3, 0.6, and 1.2 per second for 480p, 720p, and 1080p, with broad API access still forthcoming.
**AI take:** Document-to-video makes the model more production-oriented, but broad API access and audiovisual stability remain the next tests.
### Cloudflare Launches Kitesurf, a Cloud Browser for AI Agents
Cloudflare launched Kitesurf, a cloud-hosted browser designed for AI agents rather than human users. It runs common browser-automation tasks in V8 isolates, and Cloudflare says it uses less computing power than Chromium for those workloads.
**AI take:** An agent-native browser could lower automation costs, but compatibility and isolation of web behavior will determine production readiness.
### Rippling Launches a Console to Track AI Spending by Employee and Team
Rippling introduced AI Spend Console to show AI-tool spending by individual employees and teams. The product grew out of Rippling's own review after spending millions of dollars on AI within months, and aims to connect spending with usage and return.
**AI take:** AI cost governance is moving from company-wide bills to teams and tasks; the hard part is building a credible output comparison.
### Ai2 Opens TutorMoments to Test When AI Tutors Should Help or Hold Back
Ai2 released a preview of TutorMoments, which uses de-identified transcripts from real one-on-one math tutoring to test when an AI tutor should help and when it should let a student continue reasoning. The project includes code, annotated data, and scored replays; initial results suggest models told only to tutor well tend to over-help.
**AI take:** A useful tutor must choose the right moment to intervene without doing the student's thinking for them.
## Rumors
No rumor met the publication threshold in this issue.
