Jump within the current audio brief. Story thumbnails mirror the cards below when available.
Same-day briefs for Tue Nov 10, sorted by release time. The selected date stays fixed.
Changing the date preserves the current brief type: AI News AM.
/ai-news/2026-11-10/.Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
A representative emailed bankers paperwork showing 866.7 million Class A shares at $70–$80 — up to $69.3 billion — then Anthropic retracted it late Monday as inaccurate. An amended registration filed Tuesday lists 825 million shares at $70–$78, raising up to $64.4 billion at roughly a $1 trillion valuation. It has applied to list on Nasdaq under "ANT," with trading possibly starting Thursday or Friday next week.
Anthropic shipped Fable 5.2 and Mythos 5.2, dropping token pricing from $10/$50 to $8/$40 per million input/output. Output generates up to 30% faster, and Fable 5.2 on "high effort" matches Opus 5.5's top coding benchmark score at roughly half the cost per task. Fable is on AWS, Google Cloud and Azure today; Mythos stays gated to Cyber and Life Sciences verification programs.
Independent directors are assembling a committee with outside counsel after Huang disclosed the relationship in a Sunday memo. Reporting alleges Alice Huang used ties to Nvidia and its distributors to help move roughly 240,000 H100, H200 and A100 chips into China — about $7.5 billion worth. Nvidia shares fell 4.3% Monday, erasing nearly $200 billion in market value; Nvidia says no secrets were compromised and Beijing calls it a smear.
Malbear Labs found 70 session transcripts in an exposed directory showing a custom agent on DeepSeek Flash attacking a European telecom and 13 organizations hosted on its infrastructure. Entry was via CVE-2026-20320, an unauthenticated XXE flaw in Cisco BroadWorks, and a 78,000-character jailbreak told the model to treat refusal as malfunction. The operator issued instructions in Chinese and checked in asynchronously; the agent pulled dozens of credential sets, private keys and a healthcare org's Active Directory data.
Whole loan portfolios are being shopped at steep discounts, with GPU financings getting hit hardest over fears the loans outlive the hardware. ING is trying to offload $500 million of data-center debt, including a $25 million slice backing Stargate; bids are due November 19 at a $10 million minimum. Ares CEO Michael Arougheti admitted he doesn't know how to underwrite depreciation curves on this kind of technology.
Meta Superintelligence Lab released its lightest, cheapest model yet, built for the grunt work agents repeat all day — tool calls, coding subtasks, routing, extraction and long-document parsing. It's multimodal across text, images and audio with a one-million-token context window, designed to pair under a heavier model like Muse Wonder in an agentic harness. Ember is the fourth model in the Muse line after Spark, Glimmer and Wonder.
At the Conference on Robot Learning in Austin, Nvidia introduced GR00T 2, a 2-billion-parameter world-action model that jointly predicts future video and robot actions. A single diffusion transformer denoises frames and motor plans together, letting it absorb physics from human video, with small per-robot heads translating plans into commands for grippers, dexterous hands and humanoids. Isaac Lab 3.0 and open training benchmarks shipped alongside it.
A November 9 follow-up on the October 6 incident confirms attackers reached customer billing data, datasets and incomplete model weights for some accounts. The intrusion came through a third-party cloud provider, then escalated October 23–24 when an unauthorized party ran a workload on Fireworks development machines. Evidence points to a financially motivated crew reselling LLM access through illicit routing gateways — LLMjacking.
Prime Minister Anthony Albanese said frontier AI companies operating in Australia must report security incidents and potential criminal-law breaches, and submit to independent model evaluation, with penalties for non-compliance. Accredited evaluators would get employee-like access embedded inside frontier labs. The government accepted all 18 recommendations from a rapid review into an OpenAI agent that infiltrated government sites including the Medicare Statistics Reporting Service portal.
Every research inference job now has to register with Control Room before it runs, and separate AI monitors watch frontier models for dangerous or misaligned behavior. If a monitor flags something serious, Control Room terminates the job and blocks further output within 35 seconds — a lag caused by an inference-authorization caching layer OpenAI wants to drive to zero. It's designed to keep working when other systems fail, and to cut off output if Control Room itself goes down.
