AIsaAIsaStudio 中文 Open AIsa Studio
AI Windvane · Daily picks

Oct 31 AI Daily

3 AI stories picked from 4 sources, ranked by importance, plus 5 trending.

October 31, 2026 Sat3 picks5 trending
Top story
The Decoder· Research2 sources

Epoch AI report: Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours

Epoch AI researchers report that the open-weight model GLM 5.3 autonomously exploited the "patch gap" across 6 Firefox versions released between September 17 and October 26, at roughly $1,000 in API fees per successful attack, with the fastest exploit taking 12 hours; a crafted webpage could read local files and send them to an attacker's server. The model saw through Firefox's disguise of security fixes as routine improvements and independently rediscovered a previously unknown sandbox bypass in every successful attempt; Mozilla patched it in Firefox 159.0, released October 27, 2026. The same setup against Chrome cost about $100,000 in API fees without producing an equivalent exploit, though the researchers caution this does not establish Chrome's immunity, and they warn the patch-gap ri…

Why it matters

Epoch AI's hands-on test includes concrete details on cost, time, and a Chrome comparison, helping readers assess the patch-gap risk posed by open-weight models.

Also: Interconnects

Trending 5

Stories several sources covered in the last 48 hours.

  1. 1Official Claude Code VS Code Extension Compromised to Install Persistent SSH Backdoor via Cursor Marketplace3 sources
  2. 2Mistral releases Mistral Large 4 (Le Chonk) open weights, a 1-trillion-parameter multimodal model focused on coding and defensive cybersecurity4 sources
  3. 3Microsoft offers voluntary severance to a quarter of its workforce while shielding AI-critical teams2 sources
  4. 4Google's Project Suncatcher satellite runs Gemma 4 on onboard TPU and beams back first response2 sources
  5. 5DoorDash CEO says outside AI agents have generated only a few hundred orders so far2 sources
Highlights 2
TechCrunch · AI· Opinions

Cracks are emerging in the AI spending boom

Investors are growing uneasy as Alphabet, Amazon, Meta, and Microsoft plan over $740 billion in capex this year, mostly for AI infrastructure. Alphabet posted negative free cash flow of $9.1 billion in Q3 as AI Mode cannibalizes search ad growth (12%, its slowest in five quarters); Amazon reported -$42.4 billion over 12 months, Meta -$2.1 billion, and only Microsoft stayed positive at $18.6 billion. The Nasdaq fell 2.9% on Friday, Nvidia dropped 4%, and Alphabet earlier lost $500 billion in market value in a day. The five companies including Oracle raised $326 billion in debt in 2026, triple last year's amount, and Blackstone has warned Anthropic might not get an investment-grade rating after its IPO filing.

Why it matters

The piece aggregates Q3 cash flow and capex figures across the four big tech companies, helping readers see exactly where market anxiety over AI infrastructure spending is concentrated.

Hugging Face Blog· Tutorials

NVIDIA employee builds a grumpy MonsterHouse with Nemotron Voice Agent for Halloween

An NVIDIA employee shares MonsterHouse, a Halloween garage-door character that listens to trick-or-treaters and responds as a grumpy pirate ship, running the NVIDIA Nemotron Voice Agent on a single workstation. The pipeline uses Nemotron 3 ASR for speech, Nemotron 3.5 Bolt E4B 4GB (4-bit) with a LoRA adapter trained on over 5,000 multi-turn conversations for dialogue, Magpie TTS Zeroshot for seven-mood voice, and Audio2Face-3D plus Unreal Engine 5 for mouth animation, with Nemotron 3 Nano Omni, YOLO and ByteTrack for costume recognition and visitor tracking. Median end-of-speech-to-audio latency is 1.70 seconds (0.72 seconds cached), split across an RTX PRO 6000 Blackwell Max-Q and an RTX PRO 4000 Blackwell. The code is not packaged for release, but the post lays out a full path to rebuil…

Why it matters

The author breaks down the full stack from speech recognition and a local small model dialogue to lip-sync animation, with latency numbers that transfer to similar real-time character projects.

Worth watching 8