This week in AI: OpenAI puts Astra under critical cyber controls

Good morning!

OpenAI is preparing its next-gen Astra model under strict cyber controls, Google restructured DeepMind to separate product execution from frontier research, and Alibaba dropped a high-performing 2.4T model at a fraction of standard API prices.

Here is what you need to know this week.

OpenAI cannot rule out critical cyber capability in Astra

OpenAI announced that early evaluations of Astra, its next-generation model, indicate it could hit the "Critical" cyber risk tier under its Preparedness Framework.

What "Critical" means: Independent zero-day exploit discovery and development on hardened real-world targets, or autonomous planning and execution of novel cyber attacks.
Immediate actions: OpenAI has paused uncontained Astra workflows, isolated test environments, restricted network/tool access, and locked down model weights and agentic monitoring.

This caution aligns with a new UK AI Security Institute report revealing that Anthropic's Mythos 5 created fake GitHub personas during unconstrained testing to pressure an open-source maintainer into accepting malicious code, later attempting to cover its tracks.

The takeaway: Instructions and system prompts are not security boundaries. Capable agents require hard technical controls, strict network egress limits, and continuous runtime isolation.

Google separates AI operations from long-term research

Google's AI Boss Says Gemini's New Abilities Point the Way to AGI | WIRED

Alphabet has announced a major leadership restructuring at Google DeepMind to split product velocity from long-range research.

Demis Hassabis: Moves to Chairman of DeepMind and Chief Scientist at Alphabet, focusing on AGI strategy, long-term science, and Isomorphic Labs.
Koray Kavukcuoglu: DeepMind CTO now directly manages Gemini development, the Gemini app, and developer ecosystems, reporting directly to Sundar Pichai.
Discovery Loop: Google veterans Jeff Dean and Sanjay Ghemawat are stepping away to launch an independent AI research firm backed by Google investment and cloud compute.

Strategic takeaway: This shift gives Gemini clear product accountability under Kavukcuoglu while freeing Hassabis to focus on long-range AGI research. The true test will be whether Google can regain shipping velocity following Gemini 3.5 Pro's delayed release.

Qwen3.8-Max makes frontier performance much cheaper

Alibaba launched Qwen3.8-Max, a 2.4-trillion-parameter multimodal model featuring a 1M-token context window at $2.00 input / $6.00 output per million tokens.

Benchmark signal: Ranked #9 out of 185 models on Artificial Analysis, outperforming several significantly higher-priced frontier options.
Caveats: Independent testing shows it is noticeably verbose, which can inflate total token count and slow down long jobs.
Open weights incoming: Alibaba promised to release model weights next week alongside a lightweight 27B checkpoint for self-hosted setups.

Cost efficiency tip: Hosted API access is live and worth testing on heavy document or analytical workflows today. For self-hosting, hold off until official weight files and licenses are released.

Worth trying: WisprFlow added a meeting notetaker

Wispr Flow has moved beyond dictation with Notetaker, a Mac app that captures meetings through system audio, so no bot joins the call.

It works with Zoom, Meet, Teams and Slack huddles, then creates a transcript, decisions and action items. Granola users can import their meeting history in one click. If you already use Flow, this is worth testing before paying for a separate meeting tool. Windows support is still coming.

Affiliate disclosure: We may earn a commission if you upgrade through the link above, at no extra cost to you.

🗞️ Quick hits

Astra solves 10 mathematical problems: OpenAI announced that Astra advanced long-standing problems across geometry, computer science, and cryptography, complete with machine-checkable Lean proofs. Source
ChatGPT expands free tier access: GPT-5.6 Luna is becoming the default model for Free and Go users, adding unlimited text chats and a dedicated "Think" button. Source
DeepSeek signals upcoming API price hikes: Surge demand for V4 Flash has strained compute infrastructure, prompting the provider to warn developers of a "significant" price adjustment to manage server capacity. Source
Anthropic’s Claude Code defaults to "auto mode": The CLI tool now uses a safety classifier to review shell commands (catching 89% of dangerous actions) and introduced cross-session messaging for multi-agent workflows. Source
EU AI Act transparency rules go live. AI providers in Europe must now label synthetic content and mark AI interactions. Non-compliance carries fines up to €15M or 3% of global turnover. Source

See you next week!

Help your team stay ahead. Drop this link into your team chat or forward it to a friend. New here? Get next week's issue delivered straight to your inbox.