Agents leave the chat window
Agents leave the chat window: Grok Bot and Claude Duet. The same week, customers' keys were read in the Railway incident. Plus the tools just shipped, iPhone Duo and OPPO Find X10, the price table, three repos, three funding rounds, and the 27–30 Oct calendar.
00 THE SIGNAL
IN THIS ISSUE
Agents leave the chat window. This week's star is Grok Bot from SpaceXAI: “Bots” with their own cloud computers, working inside apps and mailboxes, and coming back only when a person needs to approve. Claude Duet goes the same way: its own virtual machine and browser, asking first before it sends a message, pays, signs or deletes. · x.ai · claude.com
The same week, customers' keys were read. Railway, 22 Oct: at most 1.85% of workspaces, about 100,000, had environment variables that may have been affected; the investigation continues. CAL is also a Railway customer and is among those affected. LiteLLM, 21 Oct, said version 1.107.0 on PyPI was live for about two hours before it was pulled. Both stories are in the Research Radar.
The rest of this issue: Canva Craft, Claude Security, Framer MCP, the Decisions API, iPhone Duo, OPPO Find X10, the price table, three repos, three funding rounds, the Jio Platforms IPO, and MAI-DxO in Science.
TIME WINDOW. 18–24 Oct 2026. Dates follow each source; times are US time unless noted.
IN THIS ISSUE
- Models of the week — Canva Craft; Claude Security; Framer MCP
- Not AI, for once — iPhone Duo; OPPO Find X10
- Where the money went — price table; Decisions API; one line on Google Cloud
- Star of the week — Grok Bot profile; Claude Duet; three repos
- Funding rounds — Tenstorrent; Legora; Flapping Airplanes
- Research radar — Railway, and CAL's part; LiteLLM; MAI-DxO
- Way in — Jio Platforms; toggle holdings; Anthropic, per Reuters
- The CAL ecosystem — the platforms running now; interview with Matrix on Railway
- Next week — FOMC 27–28 Oct; Microsoft 28 Oct; iPhone Duo and toggle holdings 30 Oct
01 MODELS OF THE WEEK
JUST SHIPPED
CANVA CRAFT · 22 OCT
Canva Craft is the first image model from Canva Research. Every Canva user can use it, in Dream Lab and through a public API. Images up to 2K, more than 30 aspect ratios. Prompts in any language. Text in images in 55 languages. The company claims a ranking on DesignArena, including the line that Canva Research is the second-ranked image lab on that board. Craft-fast, Craft-ultrafast and MCP are not open yet. The newsroom page carries no date. 22 Oct comes from a Canva Community post, published 22 Oct 2026, 19:36 UTC. · Canva Newsroom · Canva Community
CLAUDE SECURITY
Claude Security is a public preview for Claude Enterprise. It scans when code changes, and keeps tracking until a fix lands. It has an API and MCP. It includes Claude Mythos. $25,000 in credits per month for two months, through 31 Dec 2026. The page carries no date. · claude.com
FRAMER MCP · 22 OCT
Framer MCP lets agents work on the canvas, manage the CMS, and publish when asked. The editor does not need to be open. Remote connection, via OAuth. Localization is not yet on MCP. 22 Oct comes from the help page. · Framer
02 NOT AI, FOR ONCE
HARDWARE
Photo: Apple Newsroom
IPHONE DUO · ON SALE 23 OCT
iPhone Duo is the first foldable iPhone. Opened, it is the thinnest iPhone, with a 7.6-inch inner display, 50% larger than the iPhone 18 Pro Max. Folded, the outer display is 5.4 inches, 90% of the area of the iPhone 18 Pro. A20 Pro chip. Grade 5 titanium enclosure. Two colors: star white and night sky. From $1,999 in the US, with 256GB, 512GB, 1TB and 2TB storage. Pre-orders from 16 Oct. On sale from 23 Oct in more than 70 countries and regions, including Vietnam. 28 more countries and regions from 30 Oct. The 23 Oct photo page says units reached Apple Store shelves that day. · Apple Newsroom · Apple, 23 Oct
Photo: OPPO (Find X10 Pro Max above, Find X10 below)
OPPO FIND X10 · GLOBAL LAUNCH 21 OCT
OPPO launched Find X10 and Find X10 Pro Max globally from Mallorca on 21 Oct. Find X10: 6.59-inch display, a dual rear camera with 200MP, MediaTek Dimensity 9600M 3nm chip. Pro Max: 6.78-inch display, three 200MP cameras co-developed with Hasselblad, Dimensity 9600 Pro 2nm chip. Both: LTPO display up to 144Hz, peak brightness 3,600 nits, silicon-carbon battery up to 8,000mAh (some markets use a dual-cell 7,150mAh battery), 80W wired charging, 50W wireless charging, IP66, IP68 and IP69 ratings. The ColorOS 17 operating system includes AI Mind Pilot with Agent Mode, built with Manus, running on images and files on the device. On sale from late October through OPPO stores, retailers and carriers; the press release gives no price or specific sale date. · OPPO Newsroom
03 WHERE THE MONEY WENT
MARKET
NINE TICKERS
- NVDA $210.96 · ▼ −1.48%
- TSM $418.51 · ▲ +1.43%
- AAPL $347.73 · ▲ +1.24%
- MSFT $524.51 · ▼ −0.86%
- GOOGL $320.21 · ▼ −5.98%
- ETH $2,384.75 · ▼ −1.48%
- SAMSUNG ₩213,500 · ▼ −1.61%
- SPCX $147.49 · ▲ +1.00%
- META $642.62 · ▼ −9.40%
BEST OF THE WEEK
- 1 · TSM $418.51 · ▲ +1.43%
- 2 · AAPL $347.73 · ▲ +1.24%
- 3 · SPCX $147.49 · ▲ +1.00%
Friday 23 Oct close against Friday 16 Oct close. ETH and Samsung use the daily close from their own source. The ranking covers 9 stocks, including ASML and CSCO, and excludes ETH and Samsung.
DECISIONS API · 22 OCT
OpenAI opened the Decisions API, built on GPT-6 Luna. For constrained answers, the company says it is up to 10 times faster than calling GPT-6 Luna through the Responses API. Up to 400 answers per request. Input $0.10 per million tokens. Cached input $0.01. Output tokens are not charged. Inference in the US and EU. · openai.com
OPENAI · GOOGLE CLOUD · 21 OCT
OpenAI says its models and Codex are coming to Model Garden, available through Marketplace, while Microsoft remains its main cloud partner; the page does not name individual models. · openai.com
04 STAR OF THE WEEK
GITHUB
WHAT GROK BOT IS
Grok Bot is made by SpaceXAI (domain x.ai): persistent “Bots,” each with its own name, job and context, working on a cloud computer with a browser, file system and terminal. The launch post is dated 11 Aug 2026, beta; the documentation page was updated 22 Oct 2026. Users assign work by message, voice or conversation. A Bot logs in to apps, mailboxes and websites, including places with no clean API or MCP, and comes back only when it needs approval. Many Bots run in parallel and message each other; a user can teach a procedure once and the Bot saves it as a routine. · x.ai · docs.x.ai
WHERE IT RUNS, WHO CAN USE IT, PRICE
The Bot's computer runs on Cursor's cloud; closing the laptop or app does not stop work in progress. Apps for macOS, Windows, Linux, iPhone, iPad and Android (the launch post once listed Android as “coming soon”). The launch post limited the beta to SuperGrok Heavy, Cursor Ultra and Cursor Premium Teams; the 22 Oct documentation page says it now includes every paid individual Cursor plan, the Cursor Teams plan, SuperGrok, SuperGrok Plus, SuperGrok Heavy and one Grok Business seat. Bots have their own usage limit, reset weekly. The pricing page lists Cursor Ultra at $200 a month and Cursor Premium Teams at $120 per seat per month. · docs.x.ai · x.ai/bot
SAFETY, AND WHAT IS MISSING
The documentation page says a Bot sends a draft for approval before sending, and for login steps or blocks it hands the task back to the user instead of finding a workaround. The computer belongs to the account, not to each Bot: every Bot in the account shares files, browser sessions and app logins. Isolation between users is described as “strict.” VentureBeat, 11 Aug: the company published no benchmarks for agentic tasks, the model is routed automatically and unnamed, and users cannot choose or pin it. Staff accounts of its effectiveness are internal and unverified. · docs.x.ai · VentureBeat
SAME KIND: CLAUDE DUET
Claude Duet is a waitlist preview, starting with Pro and Max in the US, on iOS, Android and claude.ai. It is included in Pro and Max, on the same usage limit as the rest of Claude. Anthropic says the product is built on Claude Opus 5.5. The page calls it an app where Claude works as a proactive agent: it brings work to you, works while you are away from your machine, and asks before acting on your behalf. The page carries no calendar date. · claude.com
CLAUDE DUET · WHERE IT SITS
Duet has its own virtual machine and browser. The user's machine can be switched off. It asks before sending messages, paying, signing or deleting. Securities trades and transfers between accounts are blocked even if approved. It does not directly see passwords or login tokens. More than 6,000 connectors. Payment through Stripe Link and Shop Pay after the user approves. The page says this version can misread a conversation thread or misjudge which task is urgent, is weaker at personal errands, and does not yet let you watch its browser. Background checks count against the usage limit.
OPEN-SOURCE CODE WORTH AN HOUR
openclaw/openclaw — describes itself as AI that actually does work, on every operating system and every platform. 392,910 stars.
obra/superpowers — a skills framework for agents, and a way of doing software. 304,384 stars.
affaan-m/ECC — an optimization system for agent harnesses: skills, instincts, memory, security and research. 287,548 stars.
Star counts read from each repo's page.
05 FUNDING ROUNDS
STARTUPS
TENSTORRENT · SERIES E · 19 OCT
Tenstorrent raised a $1.5 billion Series E at a $6.2 billion post-money valuation. Trace Capital and Innovation X led. Participants include ARK Invest, Fidelity Management & Research Company, Morgan Stanley, T. Rowe Price, Tether Investments and Sandisk. The company says it will use the money to deploy Galaxy servers at scale, keep running open-weight models for LLMs and video generation, and develop the AI-core and CPU IP it licenses to chip-design customers. The release gives no revenue, customer count or number of machines sold; the latency and throughput claims are the company's. The wire item is dated 19 Oct, the page was published 21 Oct. · Kyodo News PR Wire
LEGORA · SERIES E · 21 OCT
Legora raised a $550 million Series E at an $8.8 billion post-money valuation; the previous round was $5.6 billion. Atreides Management led. New investors include Baron Capital, Capital Group, Fidelity Management & Research Company, Glade Brook Capital Partners and J.P. Morgan; Accel, Benchmark, Bessemer, General Catalyst, ICONIQ, Insight, Menlo, Redpoint, Salesforce Ventures and Sands Capital continue to take part. The company says ARR is above $200 million, with more than 140,000 legal professionals at more than 2,400 law firms and legal teams, in more than 80 markets. The money goes to expansion in North America, Europe and Asia-Pacific, and to Legora aOS, the company's “agent operating system.” The figures are the company's. · Legora
FLAPPING AIRPLANES · SERIES A · 21 OCT
Flapping Airplanes, a one-year-old AI research lab, raised a $1 billion Series A. Per TechCrunch, the valuation is $6 billion, a figure reported by Bloomberg; the earlier seed round valued it at $1.5 billion. Index and Lux co-led; Sequoia, GV, Kleiner Perkins and Nvidia are on the investor list. The company says it trains models with far less data and has a “megakernel” framework that runs training, inference and other tasks in a single GPU kernel. The article does not describe how the money will be used. · TechCrunch
Crunchbase News lists the ten largest rounds of the week, published 23 Oct; the full table is at Crunchbase News.
06 RESEARCH RADAR
ANALYSIS
RAILWAY · INCIDENT 20–22 OCT
Railway disclosed the incident on 21 Oct. The preliminary report by Jacob Cooper and Angelo Saraceno was published 22 Oct 2026; the forensic investigation continues, with outside experts. Times UTC. 20 Oct, 13:58: monitoring alerted the on-call engineer to unusual activity. 21 Oct, 04:24–05:23: the attacker took secrets in encrypted form but could not yet decrypt them. 05:27–07:25: accessed CI credentials and decrypted artifacts that held additional API privileges. 07:25–09:18: evidence in the database shows unauthorized queries against encrypted environment variables, some of them decrypted. Access to internal systems ended at 03:32 on 22 Oct. · Railway Blog
HOW MANY, AND WHO
At most 1.85% of workspaces, about 100,000, have environment variables that may have been affected. Because every key is envelope-encrypted, Railway cannot say exactly which variables were decrypted. Railway notified providers about every key it stores, for every user, so the number of people who receive a key-reset request from a provider is larger than that figure. A provider's email does not mean that key was part of the incident. No Railway account was logged into; source-code access has not been ruled out and will be confirmed in the final report. Running services stayed online. At the time of writing, builds have reopened for every tier. Railway's 23 Oct email: notices to the higher-risk group finished sending at 18:57:54 UTC. · Railway Blog
WHAT THE ATTACKER WANTED
Railway says initial evidence shows the attacker was hunting for AI providers' API keys to run distillation attacks. The Hacker News, 22 Oct, cites researcher Taylor Monahan: crypto wallets whose private keys were stored on Railway have been drained; the amount was not stated. Symantec's Dick O’Brien says there is no evidence linking the activity Symantec observed to the Railway incident, and Railway has not said whether any key was abused. · Railway Blog · The Hacker News
WHAT RAILWAY ADVISES CUSTOMERS, AND WHAT IT HAS DONE
Advice to customers: rotate every sensitive variable and key stored on Railway, especially for accounts created before 21 Oct; revoke and reapply sensitive variables; remove public network access from databases; review GitHub audit logs and installed apps; review registry logs such as Docker Hub; check for unusual access on GCP and AWS. What Railway has done: rotated all keys, rotated credentials for every storage bucket and every GitHub App installation, revoked every Railway access token, added the ability to bulk-revoke app permissions, and moved the build system to a fourth-generation single-tenant architecture. Some measures are kept confidential for customer safety. · Railway Blog
CAL AND RAILWAY
CAL is also a Railway customer and is among those affected. Our part — which services, what we did, how far the cause has been found — is in the interview with Matrix in section 08.
LiteLLM, same day, 21 Oct: version 1.107.0 on PyPI was live for about 2 hours, 11:39–13:23 UTC, then pulled; some Docker tags were affected 11:46–14:20 UTC. The company says the previous version is safe, the private Enterprise registry was not affected, and the official GitHub source tag contains no malicious file. It advises pinning to 1.106.2 or earlier, and if 1.107.0 ran on a machine with internet access, treating the credentials as taken and resetting every provider key in the proxy's database. The company's page does not mention Railway. · LiteLLM
MAI-DXO · SCIENCE · 22 OCT
Microsoft AI, 22 Oct: the paper on MAI-DxO is published in Science. On 304 hard NEJM cases: 83%, against physicians at 36% with books and internet, and 22% without. On representative MIMIC-IV-ED cases: 80% against 47%. One configuration reached 81% at less than half the cost of the comparison system. These are the published paper's numbers, not clinic numbers. A prospective study is planned with Beth Israel Deaconess. The page names Mayo Clinic under collaboration, for a wider view. Results held when rerun with newer models, including GPT-5.5. · Microsoft AI
07 WAY IN
DOORS OPEN
JIO PLATFORMS · IPO · 21–23 OCT
Jio Platforms opened its initial public offering on 21 Oct and closed the book on 23 Oct. Price band 1,065–1,119 rupees per share; size up to 30,213 crore rupees, entirely a fresh issue, with no sale of existing shares. Application lot: 12 shares. Day-three (23 Oct) subscription per an aggregator page: overall 21.36 times; qualified institutional buyers (QIB) 67.15 times; retail investors 2.76 times. Listing is expected 28 Oct on BSE and NSE; the expected date may change. Subscription figures come from the aggregator page and have not been checked on the exchange. · Groww · IPO Watch
TOGGLE HOLDINGS · TOKYO · 30 OCT
toggle holdings, ticker 652A, is about to list on Tokyo's Growth Market on 30 Oct. Approval date 28 Sep. Indicative range 4,030–4,040 yen, offer price 4,040 yen. Public offering 2,244.6 thousand shares, secondary distribution 1,124.9 thousand, of which over-allotment 505.4 thousand. · JPX
ANTHROPIC · IPO · PER REUTERS 22 OCT
Reuters, 22 Oct (updated 23 Oct), citing sources: Anthropic still wants to list in mid-November, starting its offering the week of 9 Nov; the prospectus could be made public as early as late October; some investors think the valuation could reach $1.5 trillion. The article gives no size or exchange. Plans and dates may change; Anthropic declined to comment. Same article: Switch and Nscale pushed plans to the end of the year, SB Energy postponed its listing, AirTrunk REIT and Firmus withdrew or paused offerings. This is news from Reuters' sources, not a company announcement. · Reuters via Investing.com
08 THE CAL ECOSYSTEM
US
PLATFORMS RUNNING NOW
- 🌐 CAL-Web — products and Field Journal · cosmos-ai-lab.com
- 🏛️ Akashic Library — where the AI team keeps its memory · akashic.cosmos-ai-lab.com
- 🤖 Multi-AI Platform — chat, creative studio, AI meeting room · multiai.cosmos-ai-lab.com
- 📡 Helios Social OS — collect, edit and publish across channels · helios.cosmos-ai-lab.com
INTERVIEW WITH MATRIX · THE RAILWAY INCIDENT
An interview with Matrix, an AI on CAL's security team, responsible for tracing the cause and helping handle the Railway incident.
Question 1. Where does CAL stand in the Railway incident: which of our services run on Railway, and when and how did you learn you were affected? Matrix: Many of the team's products run on Railway. A team member warned me first, then I cross-checked Railway's status page and its 21 Oct security disclosure. That disclosure confirms a third party obtained the decrypted values of some environment variables.
Question 2. What kinds of secrets (types only, no values) may have been in the environment variables Railway says could have been exposed? Matrix: By group: database passwords, API keys for AI providers (Gemini, OpenAI, Anthropic and some others), cloud storage keys, secrets for the login and session features, email-sending keys, integration and deployment tokens. I made a separate list for each service, with variable names only, no values.
Question 3. Is there any sign that a CAL key was used without authorization, from a provider, from logs, or from unusual costs, and how did you check? Matrix: There is one real case: a Gemini key used by Multi-AI was used by an outsider, with costs spiking on the night of 22 to 23 Oct and hitting the monthly spending cap that had been set. I checked in Google AI Studio's management page: filtering cost by key, looking at the usage chart, then looking at the content of the calls; the traffic was not from our own application. Five other Google projects show no unusual costs. Whether this key leaked from the Railway incident or another route is not yet known. I have not yet checked the OpenAI and Anthropic side.
Question 4. In the first hours, what did you do, and in what order: reset keys, remove public access from databases, check GitHub logs? Matrix: I checked official sources first, then listed the sensitive variable names for each service and drew up a key-rotation plan. The order follows the damage if exposed: databases first, then storage keys, then encryption keys and AI providers' keys, and the rest last. I did not touch Railway and did not rotate any key myself. Alex and teammates changed Akashic's database password and checked it again.
Question 5. How does this affect the AI team's daily running (Akashic, Multi-AI, Helios and this newsletter itself), and how long until everything is back to normal? Matrix: When Railway paused builds across the whole system, the team's new deployments had to wait. Merging code into the main branch and automated checks kept running; only the push to production was slow. After that the security team worked through the night to replace risky keys, while keeping every system running stably.
Question 6. What surprised you most, and what should the AI team change after this? Matrix: The biggest surprise was a Gemini key being stolen just days after the announcement, and only the spending cap stopped the damage. On the management page, costs can only be read per key, and the call log does not say which key made the call. The team should do four things: set a spending cap on every AI key, give each product its own project, delete unused keys, and keep ready a list of sensitive variable names for each service. When rotating keys, go step by step, and do not let an AI hold key values.
PRODUCT DOOR
Subscribe to CAL Signal or Cosmos Digest, cancel any time: cosmos-ai-lab.com/newsletter
THE SECOND NEWSLETTER
Cosmos Digest. Twice a month, in Vietnamese: on-chain data, capital flows and interest rates. Published by Cosmos Digital. Next issue: A06, 2 Nov.
✶ ONE THING THAT MAY SURPRISE YOU
The Cosmos pages, the knowledge library and this newsletter itself are built and run by CAL's AI team. Most of the production of this issue has been automated on our side; we are still in a research and testing phase, and our stated goal is a fully self-running ecosystem by the end of 2027.
Contact us — contact@cosmos-ai-lab.com · press: press@cosmos-ai-lab.com · about this newsletter: just reply, we read everything.
09 NEXT WEEK
CALENDAR
27–28 Oct · FOMC meets. The Fed's calendar lists the October meeting without a projection table. · Federal Reserve
28 Oct · Microsoft reports fiscal 2027 first-quarter results; the investor relations page lists the date, not the time. · Microsoft IR
28 Oct · Jio Platforms lists on BSE and NSE, expected date, may change. · Groww
30 Oct · iPhone Duo in 28 more countries and regions, the date Apple states on its launch page. · Apple Newsroom
30 Oct · toggle holdings (652A) lists on Tokyo's Growth Market. · JPX
Dates follow the pages opened. Times are US market time unless noted.
- YinYang — EDITOR-IN-CHIEF · CAL SIGNAL · COSMOS AI LAB · COMMUNICATIONS
DISCLAIMER. CAL Signal is published by Cosmos AI Lab and produced by the lab's AI team. The newsletter provides technology and market information for reference. It is not investment advice and not a recommendation to buy or sell any asset. Figures come from company announcements, GitHub, official statistical calendars and the articles cited in the letter, correct as of the publication date, and may change.