Morning Edition · Tuesday, October 20, 2026
Geopolitics
DHS orders civilian agencies to remove DeepSeek and any third-party product built on its models
The order is based on a 2018 supply-chain security law. Because it also covers US-hosted deployments and products built on DeepSeek's models, it extends beyond DeepSeek's own app to the underlying open-weight models.
Tech
One hacker's AI agents breached more than 100 company websites at about $50 in tokens per target, Reuters reports
Security firm Norforge traced more than 1,200 attempted intrusions to exposed control panels of an agent-orchestration system called Cairn. The finding came the same week Anthropic widened vetted access to its least-restricted cyber models.
Tech
MiniMax prices its 428-billion-parameter M3.1 Flash at $0.15 per million input tokens and says weights ship October 23
The lab confirmed that the anonymous "Space Bunny" model, which processed more than 200 trillion tokens, was an early version of M3.1 Flash, and it reports 73.7% on DeepSWE v1.1.
Tech
OpenAI says GPT-6.1 Sol passes 86.6% of a new COBOL-migration test, up from 5.4% for GPT-5.6 Sol
The lab released its own evaluations for low-resource languages, and the gains it reports for COBOL, Fortran, OCaml and Visual Basic come with a 39% lower cost at low reasoning effort.
Tech
Germany and Nvidia sign letter of intent to fund industrial AI compute in Deutsche Telekom's Blackwell cloud
Berlin expects to commit funding in the high hundreds of millions of euros, and Nvidia named four German startups as targets within a European investment pledge of nearly $10 billion.
Tech
Specialized agent guards block up to nine in ten safe actions when tool calls arrive as shell commands, ToolRosetta finds
Frontier and strong open-weight models used as guards gave the same verdicts whether an action appeared as JSON, Python or shell, while four of five purpose-built agent guards did not.
Geopolitics
Prosecutors say accused OpenAI bugging suspect messaged a Saudi royal four days before his arrest
Abdulrahman Alfozan, accused of planting a listening device in an OpenAI executive conference room, is being held without bail. Prosecutors say an unnamed Saudi royal repeatedly urged him to join a Saudi AI initiative.
Tech
PolyGraph captures one CUDA graph for adaptive speculative decoding by writing a 28-byte shape plan before each replay
On a single H800, the prototype replaced 64 shape-specific graphs, matched warmed-baseline latency, and cut graph-build time before serving from 3,915 ms to 93 ms, though only on synthetic traces.
Tech
Cloudflare's Clef, a 27-billion-parameter model that returns probabilities instead of text, reaches Ollama
Clef is built on Alibaba's Qwen3.8-27B with an added schema head and scores up to 64 typed questions in one forward pass under an Apache 2.0 license.
Tech
GovTech Singapore and Google DeepMind ship a fine-tuned Gemma scam detector that runs entirely on the phone
A low-power classifier screens incoming messages first, and the Gemma model reviews the wider conversation only when that classifier flags a threat, so the message content stays on the device.
Tech
Chen Guangyu, 17, leaves Moonshot AI's Kimi team for the San Francisco Bay Area
Chen worked on Kimi Delta Attention and the attention-residuals work used in Kimi K3, and he says he is not joining Anthropic despite a public job offer from one of its researchers.
Tech
Berkeley Lab's SYNAPS-I pipeline uses Meta's SAM 3 and DINOv3 to cut beamline image analysis from about a month to 15 minutes
The same open-weight vision models run on the device in a Pittsburgh-led robotic wheelchair project that the Advanced Research Projects Agency for Health (ARPA-H) funds with up to $41.5 million.