OpenAI's cyber model lands, Astra edits weights


Weekly Edition Price: Zero tokens
The Weekly Token
All the week’s news that fits the context window.
 
 
Vol. I · No. 4 · November 1 – 7, 2026
 

Hi,

Welcome to issue four of The Weekly Token. As always: the week’s biggest AI announcements, curated from X, with a link to every original post so you can go straight to the source.

This week OpenAI’s autonomous cyber pipeline patched hundreds of Linux bugs and shipped GPT-6 Cyber for vetted defenders, an MIT team showed GPT-6 Astra editing model weights by hand, Claude picked up Personal Finance and cheaper background subagents, and Moonshot opened Kimi K3.1 Pro’s weights. My picks still lean towards what matters if you’re building with agents.

I’ve given Agentic UI its own section this week, starting with Vercel’s interactive deployment card in ChatGPT, alongside plugin builders and the interfaces that bring agents into apps.

As always, hit reply and tell me what was useful, what was noise, or what I missed.

 
Lead story · Models

OpenAI on Wednesday said its security researchers built an autonomous pipeline that finds, exploits and patches vulnerabilities. Using an upcoming cyber model, a single run validated and patched 948 Linux kernel local privilege-escalation bugs for a total API cost of $1.2 million, narrowing 2,308 candidate findings down to working exploits. Thibault Sottiaux said OpenAI has assembled a cyber team and would share more.

On Thursday OpenAI introduced GPT-6 Cyber, the model behind that Linux work. It is built for security research and, OpenAI says, pairs stronger capabilities with better alignment. Security teams at Mozilla, Palo Alto Networks, Calif and Cisco, along with OpenAI's own, are already using it. Access is limited to select Daybreak Red customers for now, with a wider rollout over time.

Agentic UI
Clients that support MCP Apps get live deployment status, build logs to debug in the chat and a Visit button once the URL is ready.
Users describe an idea and turn it into a reusable plugin, from a design workspace to a sandbox game or a 3D LLM visualizer.
MCP servers that accept Claude as a trusted issuer can connect without users re-authorizing. A new enterprise-managed auth option, built on an open extension to MCP, lets admins connect plugins for a whole organization through their identity provider. Beta partners include Atlassian, Figma, Linear, Notion and Supabase.
Describe a tool and Cursor builds it, sets up a database, runs deterministic and agentic security checks, and deploys it. Vercel says the one-click deploys need no account or configuration and run on Vercel for Platforms.
CopilotKit says Antigravity file edits and shell commands can now stream live into an app, alongside generative UI including A2UI, reasoning, frontend tools, subagents and file attachments.
The Vue SDK supports CopilotChat, Sidebar and Popup, frontend tools and shared state, human-in-the-loop interactions, and generative UI using Vue components. It uses the same AG-UI events as the React SDK.
CopilotKit shared a guide on Google’s developer blog: the agent describes an interface as data, and the app renders its own components streamed live into chat.
Install with npx skills add benjitaylor/skills --skill stageshow.
Agents & MCP
DeepSeek published dsh-superagent-prototype on Saturday, an agent framework that watches a project's GitHub issues and pull requests and works through them with many agents running in parallel. Hugging Face's Leandro von Werra described it as "designed to take as little human attention as possible while delivering as much value as possible." When someone files an issue or opens a pull request, the system can take a first pass: make changes, draft a fix in its own isolated session and hand the result back for a person to review. Posts describing the repo say it listens to GitHub and Feishu hooks, iterates on feedback and only pings humans when it needs them. The code is released under the MIT license with configurable trigger rules, and DeepSeek labels it a research prototype rather than a finished product. An accompanying write-up is titled "Towards a More Autonomous Agent System."
Models
The tasks come from feedback by thousands of ChatGPT users on where they want a personal assistant's help. OpenAI says it has used LifeVal for a year to improve its models; in its results table GPT-6 Astra averages 80.6% and scores 100% on 10 of 16 tasks.
The model scores 31 on the Artificial Analysis Intelligence Index, a big jump over Inkling Small. It is available for fine-tuning and inference on Tinker now. Artificial Analysis lists it at 594B total and 17B active parameters, priced at $0.30/$1.00 per million input/output tokens, with an Apache 2.0 weights release planned.
Omni 1.2 Flash is rolling out in AI Studio, the Gemini API and the Gemini app for all Google AI users, and Runway added it the same day.
Tencent Hunyuan says Hy4 sits firmly among the leading open models and was built with cyber and biosecurity in mind. It is live on Tencent Cloud and OpenRouter, with open weights promised soon.
Tools & Products
Users connect the accounts they choose, and Claude answers money questions from their own balances and transactions. Anthropic stresses access is read-only: Claude can't move, manage or invest money. TestingCatalog reports it is rolling out to Pro, Max and Ultra users in the US.
Each running subagent refreshes its status line every 30 seconds, and each refresh used to re-read the whole conversation. Haiku 5.5 now writes it from short notes, cutting the cost of each update by about 99%.
ChatGPT Images 2.6 can make GIFs and work from user-supplied reference images. GPT-Image-2.6 also arrives in Codex and the API as Image-2.6 Flare and Sunburst, which OpenAI says generate images 50% faster; Sunburst at medium or high quality keeps details more consistent across chained edits.
Claude filtered 43M HN comments down to 6,454 videos people keep recommending. Karpathy says it took about two hours with Opus 5.5 in Claude Code, which also hardened and deployed the site. It needs no login and he shared the underlying data as a zip.
OpenAI's Tibo said account switching is live in the ChatGPT mobile apps and coming to desktop, and promised two days of banked usage resets. Both resets reached users during the week.
OpenAI says Pando cut push latency by about 80% and now serves nearly a petabyte of Git data a day, with no change to how engineers or agents use Git. Greg Brockman said OpenAI is considering offering it to others.
Security & Research
Researchers at MIT CSAIL asked GPT-6 Astra to improve Qwen3.5-0.8B's math. Without gradients or training, Astra changed 0.016% of the small model's weights — Weight-Space Programming — and posts summarizing the blog put one edit's math score at 48.1%, up from 7.9%. The post drew more than 38,000 likes. Alex Krentsel reproduced the results and said the ability is new to the latest models; Chandan Singh said Claude can write working neural-net weights from scratch.
FlashProve uses cryptographic techniques so developers can verify whether a given response came from a given AI model, and Anthropic says it is substantially more efficient than prior open-source methods. Proofs can make chat hijacks harder and let safety evaluators confirm the model under test is the one served to customers. Anthropic is open-sourcing the code with collaborators including Alessandro Chiesa, Giacomo Fenzi, Miranda Christ, Rahul Dalal and Ben Mayer.
With access to internal Anthropic data since August, METR judges that capability growth on well-specified tasks is now likely about 2x its Mar 2024–Oct 2025 pace (it would not be surprised by 1.3x–3x), and that most of the speed-up is likely driven by AI used in AI R&D. Anthropic redacted much of the evidence METR found most informative from the public report.
Industry & Open Source
Moonshot AI published the model weights for Kimi K3.1 Pro, which partners describe as its most capable 3T-class model, natively multimodal with a 1M-token context. Together AI is serving it and offering post-training on it; TestingCatalog says the weights are on Hugging Face under the Kimi K3.1 Pro License.
The fixed-term program is for academics, technical founders and staff-level engineers who want to work on AI safety. No ML background is required. Applications close Nov. 22. The post drew more than 13,000 likes.
The 65B model was pretrained from scratch, and Ricursive says it beats much larger US open-weight models such as Inkling-975B and Nemotron 3 Ultra-550B on agentic coding. The company puts its full R&D cost at $32M; a related post put Terminal-Bench 2.1 at 80.5% running on a DGX Spark.
Quick hits
■ Meta's Muse gets voice and realtime avatars in the US, and launches in Mexico — Talk by phone or video; Muse Realtime Avatar turns voice into an animated avatar.
■ Anthropic and Claude map the first 3D structure of a bacterial vault particle — Claude cut structure work from weeks or months to days.
■ Cat Wu joins OpenAI to lead product management for core products — She worked on Claude Code at Anthropic and will partner with @thsottiaux on agents that keep users in control.
■ Sarvam releases Prism, a fast decision model on Sarvam 105B — 71.8 average score at 44 ms median; API at ₹4 per 1M input tokens, ₹0 output.
■ Nous Research releases Hermes Agent v0.23.0, "The Hearth" — Desktop in a browser tab, /Rewind, group chats and spend limits for paid tools.
■ Ant Ling introduces Ling-3.2-flash, a 560B native multimodal model — 25B active; Ant Ling says it beats Opus 5 (max) on OmniDocBench1.5 and Vision2Web.
■ Meituan launches LongCat-2.5, a 1.6T multimodal model for long agentic tasks — About 48B active per token; API live with 10M free tokens for new and existing users.
■ Baidu releases ERNIE 5.5 Flash Base, a 280B open-weights MoE — 14B active, trained on 33.5T tokens, up to 1M context; vLLM added day-0 support.
■ Cognition introduces SWE-2 Flash, a fast coding model, but holds back the weights — Citing cyber risk, weights only for vetted partners; free in Devin until Nov 30.
■ Gemini API's Computer Use tool is now generally available — No breaking changes from preview; production-ready for browser, mobile and desktop agents.
■ Perplexity introduces Teammates, proactive agents for Max subscribers — They collaborate with each other and can message coworkers in Slack.
■ Google Antigravity rolls out remote agents on all paid consumer plans — Tasks run in their own cloud VMs; approve permissions from your phone.
■ OpenAI's Tibo: ChatGPT dots will run on more models, including GPT-6.1 Terra — Terra brings speed and a personality upgrade and will help bring dots to more people.
■ Elon Musk: Grok can find well-rated doctors, phone them and book an appointment
■ Jun Song says a black-hat group offered him a huge sum to post-train open models on exploits — He warns someone else will eventually take the deal.
■ Sam Altman: people should not cede control of the future to superintelligence — OpenAI rejects the idea of people being relegated to what AI systems think makes sense to reward.
■ SemiAnalysis releases GDPx, an agentic benchmark where the best model scores 12%
■ Stanford re-tested 50 widely cited LLM findings; most no longer held or weakened within 15 months
■ OpenAI plans a security sprint with Cathedral Systems and raises bug bounties, up to $172,000 for some finds
■ OpenAI announces a strategic partnership with JPMorganChase — Brockman says cybersecurity for the financial system is a focus.
 

That’s the week. If one of these changes how you build, or you think I picked the wrong stories, reply and tell me.

See you next Saturday,
Apurv

P.S. I also break this stuff down on YouTube and at apurvkhare.com.

 
Sourced from X only · every headline links to its post
The Weekly Token · Vol. I · No. 4 · Week of November 1 – 7, 2026

Apurv Khare

Frontend architecture, system design, and agentic engineering. One deep dive at a time, drawn from 10 years of building large-scale frontends and teaching 10,000+ developers. Free, no spam, ever.

Read more from Apurv Khare
Square cover of The Weekly Token newspaper, Weekly Edition, Vol. I No. 5, November 7–14, 2026, by Apurv Khare. Lead story under an Industry label: 'Jensen Huang Resigns as NVIDIA CEO', with the line 'A week-long board investigation ends with CFO Colette K

Weekly Edition Price: Zero tokens The Weekly Token All the week’s news that fits the context window. Vol. I · No. 5 · November 7 – 14, 2026 Hi, Welcome to issue five of The Weekly Token. As always, these are the week’s biggest AI announcements, curated from X, with a link to every original post so you can go straight to the source. This was one of the busiest weeks I’ve tracked. Jensen Huang resigned as NVIDIA’s CEO after a week-long board investigation. Anthropic shipped Claude Fable 5.2 and...

Square newspaper-style cover for The Weekly Token, Vol. I, No. 3, October 24–31, 2026, Weekly Edition, by Apurv Khare. The lead headline reads "Gemini 4 Carbon Takes the Top Spot," with the line "Two weeks after Argon, it leads AA Index v5 at 59 and Arena

Weekly Edition Price: Zero tokens The Weekly Token All the week’s news that fits the context window. Vol. I · No. 3 · October 24–31, 2026 Hi, Sorry this week’s Weekly Token is late. I was busy at React India in Goa. Here’s the video of my talk if you want to watch it: Watch my talk. I also ran a workshop there; it wasn’t recorded, but I’m planning to do a webinar on it soon. Welcome to issue three of The Weekly Token. As always: the week’s biggest AI announcements, curated from X, with a link...

The Weekly Token No. 2, October 17–24, 2026, by Apurv Khare. The lead story is DeepSeek's week: the open-weight V4.1-Flash-1017 and its inference stack, DeepInfer, open-sourced. Below it are Claude Duet and auto mode, Kimi K3.1 Pro, OpenAI's sandbox escap

The Weekly Token · Vol. I · No. 2 Weekly Edition Price: Zero tokens The Weekly Token All the week’s news that fits the context window. Vol. I · No. 2 · October 17–24, 2026 A note from Apurv Hi there, Welcome to issue two of The Weekly Token. As always: the week’s biggest AI announcements, curated from X, with a link to every original post so you can go straight to the source. This week was about open weights and agents that keep working after you leave: DeepSeek, Moonshot and MiniMax all...