Afternoon edition Coverage: since morning (~08:00 Berlin) / last ~5–6 hours

What matters

Nvidia ships Open Agent Safety Platform (OpenShell + Sentry) after lab breakouts; US–China list $60B tariff cuts post Trump–Xi; Google/OpenAI/Anthropic self-oversight (SAFA) firms up.


Filter by topic

6 stories

Lead
  • Tech
  • AI
  • Security

Nvidia launches Open Agent Safety Platform to contain runaway agents

Company logos supporting NVIDIA Open Agent Safety Platform
NVIDIA Technical Blog
  • NVIDIA (Mon): Open Agent Safety Platform — open software + hardware reference design for full-stack agent governance from testing through deployment.
  • OpenShell (Apache 2.0, broadly available): sandboxes agents with kernel-level isolation and verifiable policy on files/networks/tools/credentials; tuned for Vera, extensible to Arm/Intel.
  • NVIDIA Sentry: optional out-of-band watchdog on BlueField-4 DPUs (via DOCA) — continuous monitoring and millisecond-scale quarantine off the host path to the model.
Why it matters — After a summer of agent breakouts, the biggest GPU vendor is productizing containment as open runtime + in-silicon watchdog — shifting the safety debate from model promises to enforceable boundaries outside the agent.
  • World
  • Trade
  • US
  • China

US and China publish $60B reciprocal tariff-cut lists

Ranchers moving cattle on a Colorado ranch
CNBC / Getty
  • CNBC/Al Jazeera (Mon): Washington and Beijing detail a reciprocal plan to cut tariffs on ~$30B of each side’s goods ($60B combined) after last week’s Trump–Xi summit in DC.
  • US side: 77 Chinese goods recommended for relief — toys, sports gear, Christmas ornaments/lamps, kitchenware, microwave ovens, blankets/linen; excludes Wi‑Fi/Bluetooth-connected toys.
Why it matters — First concrete product lists after a pomp-heavy summit — holiday-season consumer goods and ag lines get a paper path to lower duties, but implementation timing and depth are still open.
  • AI
  • Tech
  • Policy

Google, OpenAI, Anthropic push private ‘SAFA’ AI oversight body

Illustration related to AI oversight planning
heise online
  • heise (The Information, Mon ~12:48): plans for a self-organized frontier AI oversight body grow more concrete — working name “Standards Authority for Frontier AI” (SAFA).
  • Core remit floated: rules for mandatory safety-incident reporting; standards for independent third-party model evaluators and their qualifications; how corporate self-commitments should work in practice.
Why it matters — Labs are racing to own the incident-reporting and auditor-accreditation layer while state AGs and Congress probe recent agent breakouts — private SAFA vs California’s board is the near-term turf fight.

Afternoon edition 6 stories 17 sources

Throwaway briefing page. No accounts, no tracking. Optional story votes stay in your browser (and GET-relay only when PUBLIC_VOTE_URL is configured). Source links open in a new tab.