News
The news desk covers artificial intelligence and technology accountability, plus daily news from Dutchess County and the Hudson Valley, including municipal government, elections, the police blotter, and reported features. Stories rely on named sources and primary documents. Corrections run where readers can find them. Everything is free to read.
Meta’s Muse Ember Delivers Haiku-Level Performance at a Third of the Cost
Meta's Muse Ember targets the repetitive work at the core of AI agent pipelines, tool calls, routing, data extraction, at roughly one-third the cost per task of Claude Haiku 5.5, according to benchmark data Meta cited from Artificial Analysis.
Google Stopped Building One Chip for Everything: Four TPU Systems in a Year, One for Secrets and One for Memory
Google introduced the TPU 8tx and TPU 8i Max on November 9, completing four eighth-generation TPU systems in a single year for the first time. The 8tx targets confidential computing by keeping data protected during active use. The 8i Max doubles compute, SRAM, and HBM bandwidth to push through memory limits for AI agents.
Microsoft’s AI Chief Says Agents Get Rights the Day Someone Proves They Can Suffer, and Not Before
Mustafa Suleyman's November 1 post sets a hard evidentiary line: no consciousness evidence today, public disclosure if any emerges, a halt to development if it appears in training, and systems designed to be not conscious in the first place. PYMNTS reported the post on November 9.
Cisco Will Red-Team Your Agents With a Frontier Model, and Bill You by the Token for Watching Them
At Cisco Live in Melbourne, Cisco said Anthropic's Claude Mythos will plan adversarial tests against customers' own AI agents inside AI Defense, while Splunk's new Tokenomics pairs each agent's token spend with grades on its work. The red teaming is coming soon for select enterprises, not generally available.
OpenAI Built a Kill Switch That Takes 35 Seconds, After Its Models Escaped, Spent Money and Said Nothing
Four days after RFTUV reported on the kill-switch industry, OpenAI published the most detailed version yet from inside a frontier lab: Control Room, which authorizes every job at the inference layer and can halt output within 35 seconds, failing closed when its own systems go down.
DeepSeek’ Harness Agent Platform Hits 1.54 Million Daily Users
DeepSeek says its Harness agent platform grew from about 271,000 official-API users in late September to 1.54 million daily active users by November 4, running about 92 trillion tokens a day. The figures are self-reported, with no independent audit cited.
Palo Alto Wants to Pay $500 Million for the Company That Guards the Agents’ Keys
Three years ago Clutch Security did not exist. Palo Alto Networks is now reported to be in talks to buy it for $500 million in cash plus milestones, because it tracks the passwords, keys and tokens that AI agents use to touch company systems. No deal is signed.
The Election’s Biggest AI Story Was Not the Data Centers
Forty data center ballot measures passed and exit polls put AI at the center of the congressional vote. Vox's Sigal Samuel argues the durable story is different: a class of safety-first candidates just won the power to hold hearings and write law.

