OpenAI Flags Astra As Its First-Ever Critical Cybersecurity Risk Model

Today's AI news: OpenAI classified its upcoming Astra model as its first ever critical-risk model for cybersecurity and is slowing its release to add safeguards, Claude Code sessions can now message each other directly, OpenAI launched an open Agent Plugins standard with AWS, Cursor, GitHub and Vercel, and HyperFrames introduced an MCP that turns Claude Design prototypes into polished video. Chapters: 0:00 Today's AI News 0:35 OpenAI Astra Critical Model 1:26 Claude Code Cross-Session Messaging 2:01 OpenAI Agent Plugins 2:38 HyperFrames MCP Video Blend Roundup 2026-08-07 https://x.com/OpenAI/status/2085801349866729975 https://x.com/ClaudeDevs/status/2085817074816070014 https://x.com/OpenAIDevs/status/2085398373511918022 https://x.com/HyperFrames_/status/2085788866691416065 [quick] OpenAI classified its upcoming Astra model as its first ever critical-risk model for cybersecurity and is slowing its release to add safeguards, Claude Code sessions can now message each other directly, OpenAI launched an open Agent Plugins standard with Cursor, GitHub, AWS and Vercel, and HyperFrames introduced an MCP that turns Claude Design prototypes into polished video. [excited] Let's get into today's AI news. [quick] OpenAI has classified its upcoming model, Astra, as the first model ever to reach the critical threshold for cybersecurity risk under its Preparedness Framework. Internal evaluations over the past few days showed Astra making significant advances in agentic coding and cybersecurity, to the point where OpenAI can no longer rule out it having critical cyber capabilities, meaning identifying and exploiting hardened real-world systems largely without human help. In response, OpenAI is slowing Astra's development, isolating its test environments, restricting its network and tool access, hardening how its weights are stored, and monitoring every agentic run for risky behaviour, with government agencies and safety groups helping test it. It's the first time any lab has actually triggered its own critical-risk threshold rather than just planning for the possibility, and it's a genuinely significant marker for how fast frontier models are approaching capabilities that were previously treated as theoretical. [quick] Claude Code sessions can now message each other directly. Instead of re-explaining your context every time you switch to a different session, you can tell Claude to send a summary to another one, and that session picks the work up mid-task. It sends a summary rather than your full history or files, so you're not exposing everything from one session to another, just enough for it to continue coherently. For anyone juggling multiple parallel Claude Code sessions on related work, this removes a genuinely annoying bit of manual context-copying. [quick] OpenAI has introduced Agent Plugins, an open standard for packaging agent skills and MCP server configurations so you can build a plugin once and use it across multiple compatible agent clients. It was developed jointly with AWS, Cursor, GitHub, and Vercel, meaning it's already got real cross-vendor backing rather than being a single company's proposal. Right now, building for multiple agent platforms usually means redoing the same integration work several times over in slightly different formats, and a shared standard like this could meaningfully cut down on that duplicated effort. [quick] HyperFrames has introduced an MCP for Claude Design that turns an animated prototype into a finished video. You animate something in Claude Design, hit share, and the HyperFrames agent scores it, adds sound effects, tightens up the motion, and renders it out as an MP4. It's a neat example of chaining one AI tool's output directly into another's pipeline, turning a design prototype into a shareable video without you needing to touch a separate editing tool at all.