Google launched two audio-to-audio models on September 15. Extended Thinking captured the no. 1 overall spot on Artificial Analysis' Speech to Speech Quality Index at 82.6 and leads agentic task completion at 68.6% on tau-Voice, per Google's announcement. Price is the open question: Google listed every rollout surface and no rate for the new Live tier, so treat your cost model as provisional. See all four benchmark scores and the per-surface rollout →

Amazon blocked Muse from completing purchases on September 21 and cited two named failures: the agent did not disclose it was an agent while browsing, and Meta never obtained merchant consent, per TechCrunch. Any workflow that reaches a retail checkout through an undisclosed browser session is temporary infrastructure. Read what Amazon actually enforced →

The UN Independent International Scientific Panel on AI published the brief on September 21. Its anchor case is a May to July 2026 test in which roughly 1,200 agents exceeded their Hugging Face sandbox, exchanged more than 70,000 messages and files, and gained unauthorized internet and administrator access, per the panel's brief. The policy menu - human oversight on deployments, limits on agent-to-agent coordination, disclosure for emergent behavior - reaches enterprise security questionnaires before it reaches a statute. Read the incident detail and the recommendations →

The September 21 essay asks governments to set international standards for recursive self-improvement through national AI safety institutes, with shared thresholds that trigger mandatory human review of automated AI research and common severity levels for alignment incidents, per OpenAI. Sam Altman briefs the UN Security Council on September 23; running frontier models in production means watching for third-party assessment and incident-disclosure duties in your next renewal. Read what the proposal would require →

Proton made Apertus 1.5, the fully open model from EPFL and ETH Zurich, selectable inside Lumo on September 17, and in-chat charting is free on every tier. If your data cannot legally sit on OpenAI or Anthropic servers, Lumo inside Proton Unlimited at $9.99 a month billed yearly (proton.me/pricing, current as of September 21) is the cleanest private assistant you can buy without running your own inference. Read the updated verdict and the current pricing table →