Skip to content
DigIntoAI/Model intelligence
Model intelligence feed

Model intelligence.

Follow the changes. Understand what comes next.

New model releases

Editor’s pickGoogle
Google/Gemini 4 Neon

Google Launches Gemini 4 Neon in Antigravity With 1M Context, Four Thinking Levels, and Benchmarks vs. Claude Opus 5.5 and GPT-6.2 Sol

Announced on November 19, 2026 at 05:51 Beijing time, Google's Gemini 4 Neon features a 1M-token context window, four thinking levels, and planner-executor pairing with Gemini 4 Hydrogen in Google Antigravity, outperforming Claude Opus 5.5 and GPT-6.2 Sol across most of 18 published benchmarks.

Latest stories508

Newest published
01

NebuSec Breaks Down Meta Muse's 3-Stage Sandbox Bypass (Patched): Silent Sentinel Egress & Shared-Kernel Lessons

On November 20, 2026, NebuSec published technical research detailing how a single malicious webpage could achieve 0.5-click RCE in Meta Muse's bundled PDF-export Chromium, silently exfiltrate private chats and memories past the Sentinel egress gate via bodiless GET/HEAD/OPTIONS requests, and escape a shared Linux 7.0 kernel to per-user VM root. Meta patched every exploit before publication, leaving four concrete lessons for enterprise AI agent sandboxes.

Meta Muse; SentinelRead story
02

AWS Threat Intelligence: Autonomous Cairn AI Agents Found and Exploited a MONAI Label Zero-Day in 11 Minutes

On November 20, 2026, Amazon Threat Intelligence disclosed a real-world campaign in which autonomous Cairn AI agents—given only the goal 'obtain shell access'—discovered and exploited a path traversal zero-day (GHSA-qc22-rmgc-vg94) in open-source medical imaging platform MONAI Label (0.8.5 and earlier) in roughly 11 minutes to obtain root access on a university host. Although a code fix was merged on October 19, no patched official release has shipped yet.

Amazon Threat Intelligence; Cairn; MONAI LabelRead story
03

Moonshot AI Launches High-Speed Kimi K3.1 Pro Prime on Alibaba Model Studio: 1.5x–2x TPS with 1M Context Intact

Listed on November 19, 2026 on Alibaba Cloud Model Studio's Prime Mode (and tracked across global gateways on November 20), kimi-k3.1-pro-prime is the high-speed Prime variant of Moonshot AI's 3T-class Kimi K3.1 Pro. It keeps identical reasoning, vision, tool-calling, and 1M-token capabilities while boosting output TPS by 1.5x–2x at RMB 26 input / RMB 136 output / RMB 1.3 cached hit per 1M tokens in Beijing.

Kimi K3.1 Pro Prime; Alibaba Cloud Model StudioRead story
04

Together AI Discloses Prompt Sample Leak After AI Coding Agent Exposed Private Credential During xorl Development

On November 21, 2026, Together AI disclosed that an unauthorized party downloaded customer-authorized serverless testing prompt samples (primarily MiniMax M3, GLM-5.2, and Kimi K2.6 requests from June–July 2026) after an AI coding agent committed a private Hugging Face repository credential while developing its public xorl GitHub repository. The intruder actively harvested and tested secrets inside the prompts—including five Dragos canary tokens—prompting Together AI to make ZDR the default for new accounts.

Together AI; xorlRead story
05

Why AI Router ZDR Labels Are Not a Guarantee: Inside OpenRouter's MiniMax Warning and a 4-Step Hardening Guide

Following Dragos's November 19, 2026 study of two million canary-seeded prompts, OpenRouter emailed affected customers on November 20–21 warning that prompts and outputs sent to MiniMax models from July to October 2026 may have been read by a third party—whether served directly or forwarded via AtlasCloud, Novita, or StreamLake—and removed ZDR labels from affected endpoints pending investigation.

OpenRouter; MiniMaxRead story
06

ChatGPT Space Rolls Out Globally to Plus Users Ages 18+, Separate from Dot's Regional Limits and Slides Rollout

On November 20, 2026, OpenAI announced in its ChatGPT release notes that ChatGPT Space is rolling out globally to Plus subscribers ages 18 and older. Independent from the previous day's region-restricted Dot test, Space carries no EEA/UK/Switzerland exclusion for Plus—while Slides rolls out separately to eligible Space accounts, meaning Space access does not automatically unlock Slides.

ChatGPT Space; ChatGPT SlidesRead story
07

Alibaba CEO Eddie Wu Discloses Next-Gen Flagship Qwen Training on In-House T-Head Chips as U.S. Compute Curbs Tighten

On Alibaba's Q2 FY2027 earnings call on November 19, 2026, CEO Eddie Wu disclosed that Alibaba has begun training its next-generation flagship Qwen model on proprietary chips designed by its T-Head semiconductor unit, projecting that after 2027 the vast majority of new AI chip supply will come from T-Head silicon amid new U.S. restrictions on overseas AI compute rentals.

Qwen; T-Head Zhenwu V900Read story
08

Perplexity Adds Decider Prompt-Injection Screening to Context API: Practical Guide to Safe Web Fetching and Nov API Stack

On November 19, 2026, Perplexity integrated its Decider decision model into the newly launched Context API ($1 per 1,000 fetched URLs, failed fetches free), letting developers run yes/no screening questions scored from 0 to 1 to catch prompt injection before web content enters an agent's context. This guide covers Perplexity's Nov 17–19 API suite—including zero-markup Orchestrator API—and a four-step production defense checklist.

Perplexity Context API; Decider; Orchestrator APIRead story
09

Inside OpenAI's Performance Factory: Cloud Codex Agents Track 1,700+ ChatGPT Investigations and Root-Cause 600+ Regressions

On November 19, 2026, OpenAI detailed its internal Performance Factory—an autonomous cloud Codex system built around a configurable task service, dispatch loop, scoped cloud Codex API, and event subscriptions. Over three-plus months, its Regression Workflow tracked 1,700+ investigations (approx. 17/day) and root-caused 600+ regressions using adversarial peer-review turns while keeping human engineers in charge of shipping fixes.

Performance Factory; OpenAI CodexRead story
10

Thinking Machines Lab Open-Sources Inkling 2 Mini: A 26.4B Looped MoE Model Running Offline in 20 GB

On November 20, 2026, Thinking Machines Lab released Inkling 2 Mini under the Apache 2.0 license—a multimodal looped MoE transformer (26.4B total, 3.4B base active, 5.4B effective active across 3 middle-layer loops) with a 1M context window. Trained and deployed natively in mixed NVFP4 and MXFP8 (approx. 20 GB single checkpoint), it runs locally on 32 GB+ Apple Silicon Macs or a single NVIDIA DGX Spark while leading sub-100B open-weight coding benchmarks.

Inkling 2 Mini; TinkerRead story
11

OpenAI Details Anti-Distillation Defenses and GPT-6.2 API Safeguards: Reasoning Extraction Drops to 0.2%

Published November 20, 2026 (promoted by @OpenAI on November 21 at 03:33 Asia/Shanghai), OpenAI's research post 'Combatting adversarial distillation' examines how exact reasoning traces transfer dual-use cyber capabilities and outlines GPT-6.2 Astra and Sol's upgraded request- and campaign-level defenses—cutting adaptive black-box extraction success from 54.0% to 0.2% while binding API reasoning to exact conversation histories.

OpenAI APIRead story
12

Offline AI Inside Microsoft Edge Web Pages—With a Catch: Aion Instruct 2.3B Requirements, Origin Trial, and Offline Demo Guide

On November 19, 2026, Microsoft replaced Phi-4-mini in Microsoft Edge with the 2.3B-parameter Aion Instruct (MIT-licensed on Hugging Face) and opened a Prompt API Origin Trial for live websites. Once downloaded, inference runs locally offline with no cloud API, but PCs must meet strict gates: 3 GB VRAM, 20 GB free disk space on the Edge profile volume (not the model size), and an unmetered network for initial download.

Microsoft Edge; Prompt APIRead story

From the Webmaster

Personal experiences, thoughts and notes.

View all notes