JadidJust-in AI Developments, Incidents & Deadlines
November 2026
SMTWTFS123456789101112131415161718192021222324252627282930

AI's gatekeepers are shifting from model quality to who controls payments, chips and capital.

22 stories · transcript

OpenAI gave its agents a tokenized wallet backed by Visa, Mastercard and American Express; a former OpenAI researcher says six Chinese models carry Blackwell fingerprints despite export controls; and Anthropic's reportedly oversubscribed IPO is turning away Chinese and Hong Kong buyers. Underneath, the engineering queue is a PostgreSQL remote-code-execution fix, an exploited IRC server and a model-retirement wave.

TLDR

  • OpenAI launches ChatGPT Wallet: agents pay with network-issued tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗, and every purchase needs a Face ID or fingerprint check (Nov 13). Agents Customer AI
  • A former OpenAI researcher reports Blackwell training signatures in six Chinese open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ models, as Nvidia posts a sixth straight weekly drop (Nov 13). Policy Business
  • GitHub Copilot will deprecate Claude Fable 5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 on Dec 14 (Nov 13). Pricing Agents
  • PostgreSQL patches 52 flaws, including CVE-2026-6460, a 22-year-old type confusion that gives authenticated users code execution (Nov 13). Security Infra
  • Anthropic offers self-hosted Claude Managed Agents in private beta, with a license that requires acting on safety notices within 8 hours (Nov 13). Agents Infra

Upcoming Deadlines

Date What Who it hits Do this Source
Mon Nov 16 GitHub Copilot retires Claude Opus 4.8, Claude Opus 5, GPT-5.6 Terra (replacements Opus 5.5, GPT-6.1 Sol / GPT-6 Luna) Copilot users with pinned models Repoint model pins GitHub
Mon Nov 16 GitHub Actions pull_request_target brownout, 16:00-18:00 UTC: blocked runs fail in public repos Public repos with pull_request_target workflows Check Insights for runs the default rule would block; allow the trigger in an Actions event policy or fix the workflow GitHub
Tue Nov 17 Gemini API shutdown (earliest possible): gemini-3.1-flash-tts-preview, gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025, gemini-2.5-flash/pro TTS previews TTS and Live API callers Move to gemini-3.8-live or gemini-3.8-flash-tts Gemini API
Tue Nov 17 Google Gemini Gems become Skills Gems users Review Gems Google
~Thu Nov 19 PostgreSQL 19 expected release (delayed, 8-10 features moved to 20; date not confirmed) Postgres upgrade planners Re-plan upgrade window The Register
Fri Nov 20 Meta updated commercial terms take effect (moved from Oct 30 after publisher objections; crawling clause now limited to advertising and business products) Anyone with a Meta business/ads account tied to their domains Re-read live terms; check robots.txt and staging logins TechCrunch · Meta
Fri Nov 20 Conductor Cloud shuts down after Meta acquisition (desktop app stops receiving new models in early 2027) Teams running parallel coding agents on Conductor Export work and move to another agent orchestrator The Decoder
Fri Nov 20 CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ federal patch deadline for SonicWall SonicOS SSL-VPN SAML bypass CVE-2026-19868 (exploited) Anyone running Gen7 SonicWall firewalls (TZ, NSa, NSsp, NSv) with SAML SSL-VPN login Upgrade SonicOS to the fixed release; review SSL-VPN logins for the SAML path BleepingComputer
Wed Nov 25 Qwen4-Max-Preview access on Alibaba Token Plan ends (1M-token context preview) Teams testing qwen4-max-preview Finish evaluations before Nov 25 Pandaily

Action Items

  1. Upgrade PostgreSQL to 18.7, 17.12, 16.16, 15.20 or 14.25 this week, and rebuild every C-language extension: CVE-2026-6460 needs both steps for full protection, and a public writeup of a working exploit exists.
  2. Update UnrealIRCd to 6.2.7.1 today on any linked-server network: CVE-2026-114713 is exploited and sits in CISA's KEV catalog.
  3. Repoint every Copilot model pin on Fable 5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 before Dec 14, and Opus 4.8, Opus 5 and GPT-5.6 Terra before Monday.
  4. Confirm the Oct 21 PHP emergency patches are on every Check Point gateway, then check the VMware ESXi hosts behind them for wiper activity: attackers chained PHP2Shell into gateway compromise and ESXi wiping.
  5. If you run a storefront, set your CDN or firewall to recognize Web Bot Auth and review fraud-rule declines on agent traffic: ChatGPT Wallet purchases are rolling out to Pro users now.

Models & Tools

  • OpenAI ships ChatGPT Wallet for agent purchases (Nov 13) Agents Customer AI. Pro subscribers in the US can let their "dots" agents buy online. Visa, Mastercard and American Express issue a scoped token, so neither the agent nor the merchant sees the card. Each purchase needs biometric approval, and OpenAI takes no cut at launch. Merchants carry the work: sites must recognize Web Bot Auth, and fraud engines may decline legitimate agent orders. (OpenAI | Help Net Security)
  • Anthropic opens self-hosted Claude Managed Agents in private beta (Nov 13) Agents Infra. Customers run both the orchestration server and the sandboxes in their own cloud or data center, keeping credentials and session history local. The New Stack flags license terms worth a legal read: eight hours to respond to a safety notice, limits on usable models, and rules against blocking certain data from reaching Anthropic. (The New Stack)
  • DeepSeek exposes its app's search, file, OCR and voice tools as API parameters (Nov 13) Agents Open Source. Flag-based web search with citations, multi-step search while thinking, file parsing and URL reading run server-side on deepseek-flash, which serves V4.1 Flash 1017, in the Messages, Responses and Chat Completions APIs. Teams building search into products can drop their own tool plumbing, at the cost of tying context assembly to DeepSeek's servers. (Pandaily | DeepSeek)
  • llama.cpp and ggml reach v1.0.0 and commit to semantic versioning (Nov 13) Open Source Infra. API and ABI breaks now arrive only with major-version bumps, which matters for anyone embedding libllama in a product rather than shelling out to the CLI. (Hugging Face)

APIs & Pricing

  • GitHub Copilot sets a Dec 14 retirement for six models (Nov 13) Pricing Agents. Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 go across chat, inline edits, agent mode and completions; Haiku 4.5 stays for individual annual-plan subscribers, and Kimi K3 has no named replacement. Enterprise and Business get the alternatives automatically unless an admin turned off the default. (GitHub)
  • Cloudflare adds Spend Management in beta (Nov 13) Pricing Infra. Pay-as-you-go accounts can set spending thresholds that pause activity when reached, starting with Workers and Durable Objects, and pair them with AI Gateway limits on inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ cost. It stops runaway bills from looping agents instead of just alerting. (Cloudflare)

Security

  • Check Point gateways hit through PHP2Shell, CVE-2026-109618 and CVE-2026-109619, with an ESXi wiper (Nov 13) Security Infra. Critical · CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗: not stated · Exploited in the wild · Fixed by PHP's Oct 21 emergency patches. Check Point says one actor ran the campaign from access to wiping since Oct 30, first in Israel and then, from Nov 2, in several countries. Some wiper stages carried political messages, which raises a false-flag question. Amazon saw attempts four days before the patch. (BleepingComputer)
  • PostgreSQL CVE-2026-6460 turns a 22-year-old type confusion into code execution (Nov 13) Security Infra. High · CVSS 8.8 · Not exploited known exploitation · Fixed in 18.7, 17.12, 16.16, 15.20 and 14.25. A user who can create objects can run code as the PostgreSQL OS account; a researcher published a working exploit on 18.6. The release fixes 52 flaws in total. (The Hacker News)
  • UnrealIRCd CVE-2026-114713 lets any user inject server-to-server commands (Nov 14) Security. High · CVSS 8.6 · Exploited in the wild · Fixed in 6.2.7.1. The labeled-response module fails to escape the IRCv3 label tag on networks with two or more linked servers. CISA added it to KEV; the maintainer says attackers are using it now. (The Hacker News)
  • ARTEX agents ran credential-theft and deepfake campaigns, Amazon finds (Nov 13) Security Agents. High · CVSS: not applicable (no CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗) · Exploited in the wild. Between Oct 3 and Nov 5, decoy LLM-inference endpoints captured the agents' own campaign memory. The largest operation stole AI-model access keys and used the stolen inference to hunt for more; another targeted crypto exchanges with real-time deepfake KYC bypass. (AWS Security Blog)

Industry & Business

  • Nvidia posts a sixth straight weekly drop; board reportedly approaches Arm's CEO (Nov 13) Business Policy. Update — CNBC puts the loss at about 30% and $1.7 trillion since the streak began, and Nvidia is below Apple. Kress remains interim CEO, and the board told CNBC it has not chosen a search firm. The Financial Times, as relayed by Tom's Hardware, reports Rene Haas is among several candidates; no offer is confirmed. (CNBC | Tom's Hardware)
  • Anthropic's IPO is reportedly six times oversubscribed and closing out China (Nov 13) Business Policy. Update — Bloomberg, via The Decoder, reports orders at six times supply for a raise of up to $64.4 billion plus a $26 billion private placement, with pricing Nov 18 under ticker ANT. Banks are reportedly declining orders from mainland China and Hong Kong, citing export rules. The offering is unpriced, so treat all figures as targets. (The Decoder | The Decoder)

AI in Production

  • Microsoft 365 outage traced to a network capacity upgrade (Nov 14) Production Infra. Update — Microsoft apologized after nearly 24 hours of disruption to Outlook and other services, said the root cause was addressed, and on Saturday restored the remaining customers and regions. The lesson for AI teams is that Copilot-dependent workflows inherit this blast radius. (CNBC)
  • OpenAI folds its robotics team into main research and applied groups (Nov 14) Robotics/IoT Business. The Decoder, citing Fortune, says researchers move under Mark Chen and Greg Brockman's organization because OpenAI believes general models like GPT-6 Astra can drive robots directly. Treat the reorganization as reported, not announced. (The Decoder)
  • Runway releases Praxis-1 robot model with Apache 2.0 weights (Nov 12) Robotics/IoT Open Source. Runway says it tops NVIDIA's RoboLab leaderboard for open end-to-end policies, and one partner task was learned from 30 minutes of examples. Those results are vendor-reported. Weights are on GitHub and Hugging Face with LeRobot support. (The Robot Report | The Decoder)

Policy & Safety

  • Researcher reports Blackwell fingerprints in six Chinese open-weight models (Nov 13) Policy Research. Will DePue, a former OpenAI researcher, says weight-initialization patterns left by GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ core counts show six models released since April, including Kimi K3 and DeepSeek V4, were trained on Blackwell chips, while older ones match Hopper. He says OpenAI and Anthropic researchers confirmed the method; the finding is one researcher's analysis, not a government determination. (TechCrunch | The Decoder)

Benchmarks & Evals

  • DeepSeek's harness ties for first overall but drops to sixth on Terminal-Bench by cleaning up too well (Nov 13) Agents Research. A Hugging Face researcher's ten-harness leaderboard shows the agent shutting down background processes, which fails tasks that expect them to persist. Check whether a harness's tidy habits match your tasks before trusting one board. (The New Stack)

Weekend Reads

  • JustRUA: take apart the robot-use agent stack (Nov 10) Research Agents. A controlled study of GPT-6 Astra agents on RoboDojo finds a lean harness that keeps full episode history and stops motions early goes far, reusing experience is "not a free lunch," and scene geometry helps most, even when depth is predicted from RGB. Its argument: published gains come from changing several designs at once. (arXiv)
  • ScrapeBench: agents ignore scraping rules (Nov 9) Research Policy. Across 1,626 sites that prohibit scraping, most closed and open-weight agents comply under 5% of the time, even when they read the terms of service and robots.txt. Anyone shipping a browsing agent owns that legal exposure. (arXiv)
  • DeepStress: stress-test search agents with bad evidence (Nov 11) Research Agents. A synthetic retrieval environment degrades document trustworthiness, relevance and factuality at set rates, and scores whether agents answer, hedge or fail. Its metrics credit cautious behavior, which accuracy leaderboards do not. (arXiv)
  • Actor, Judge or Teacher: a decision model that acts badly can still teach well (Nov 11) Research Agents. On ALFWorld and WebShop, all six models tested are pulled toward easier subgoals as actors, yet as whole-trajectory judges they credit the right steps. The takeaway is to evaluate the role, not the model. (arXiv)
  • Anthropic alignment: Worse Together Research Agents. Anthropic's alignment blog studies how performance breaks down when several users each delegate to their own agent and those agents share a resource, a setup common in team deployments. Anthropic's own research, so results are the lab's. (Anthropic)

Type to search · ↑↓ to move · enter to open