AI Daily Brief · Oct 19, 2026 | Musk pledges a SpaceX share to every US child; one hacker's AI agents breach 100+ company sites; MiniMax launches M3.1 Flash

The most-engaged post in AI over the past 24 hours came from Elon Musk, who picked up on Anthropic’s new Public Benefit Fund and pledged to give one of his SpaceX shares to every child in America next year. On security, Reuters revealed that a single hacker broke into more than 100 companies’ websites while leaving almost all of the work to AI agents, and Microsoft expanded a system that treats agents as potential insiders. MiniMax released M3.1 Flash and revealed that Space Bunny, the stealth model that topped OpenRouter for 10 straight days, was an early version of it. Wall Street avoided a “Black Monday”: AI chip stocks bounced at the open after last week’s rout, while developers argued over Codex versus Claude Code and how far to trust AI outside their own expertise.

Today’s highlights

• Musk says he will give one of his SpaceX shares to every US child next year — about $10.5 billion by CNBC’s math (roughly 72 million children at Friday’s closing price) — but has given no details.
• Reuters exclusive: a Chinese-speaking hacker’s AI agents tried more than 1,200 websites and breached over 100 companies, including Deere, SharkNinja and Cummins, for roughly $5,000 in tokens.
• MiniMax launches M3.1 Flash at $0.15/$0.60 per million input/output tokens, with weights due Oct 23; OpenRouter’s Space Bunny was an early build.
• US stocks open Monday with the Nasdaq up about 1% and the S&P 500 up 0.5% as chip stocks rebound; the Philadelphia semiconductor index fell 11.4% last week.

1. Musk answers Anthropic: one of his SpaceX shares for every US child next year, about $10.5 billion at Friday’s close

Anthropic had announced a Public Benefit Fund — 5% of its shares up front, subject to shareholder approval, with the board able to add another 2% for every 50% rise in valuation above a baseline — and Musk had replied that he recommends giving shares directly to the public and that this is what he will be doing. On Oct 19 he went further.

The pledge: [Personal pledge by Musk; no details announced] “Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many,” Musk posted. With 327K likes and 12.77M views, it was the most-engaged post of the past 24 hours. CNBC counts about 72 million Americans under 18 (Census, July 1, 2025); at SpaceX’s Friday close of $146.03 that comes to about $10.5 billion. Musk holds about 4.8 billion SpaceX shares, or 6.4 billion including unvested awards and options (a 48.4% beneficial stake). He did not say how the giveaway would work, and CNBC said it had asked SpaceX for comment. Our own check: 72 million × $146.03 ≈ $10.51 billion, and 72 million ÷ 6.4 billion ≈ 1.1%, in line with the 1.12% calculated by @kimmonismus.

Anthropic’s numbers: [Media report] The Financial Times puts Anthropic’s valuation at $965 billion, making the 5% stake worth about $48 billion (965 × 5% ≈ $48.25 billion); the baseline has still not been disclosed. The BBC notes that Anthropic is preparing for a listing that could value it at more than $1 trillion, and that SpaceX’s IPO in June raised a record $85.7 billion.

Reactions: @kimmonismus said Musk had just confirmed “the beginning of Universal High income” (3,332 likes, 700K views); @IterIntellectus called it “the beginning of universal basic income” (3,551 likes); Gary Marcus pushed back that what Musk said isn’t binding and gives no amount or date (251 likes). These are personal views. Anthropic’s blog speaks of possible “direct distributions to specific populations” and does not mention UBI.

Sources: @elonmusk · @FT · @kimmonismus 1 · @kimmonismus 2 · @IterIntellectus · @GaryMarcus · CNBC · BBC · AP · Anthropic blog

2. Reuters exclusive: a Chinese-speaking hacker let AI agents do the work, breaching 100+ company websites at about $50 in tokens per hit

[Media exclusive] Reuters reported on Oct 19 that Snorre Fagerland, principal threat researcher at security firm NorForge, has been scanning the web for accidentally exposed instances of Cairn, an open-source AI “problem-solving” platform popular with hackers. Starting around Oct 2 he found about two dozen being used maliciously, which let him watch over attackers’ shoulders as their AI worked. The most prolific was a Chinese-speaking hacker who, from Sept 25, used instructions of fewer than 100 Chinese characters to set agents loose on e-commerce sites in the US, Canada, Australia and Japan. By late last week the agents had attempted more than 1,200 sites and breached websites belonging to more than 100 companies, with a successful intrusion typically taking about 23 hours. Fagerland estimates the campaign used roughly $5,000 worth of tokens — about $50 per successful hack — apparently mixing DeepSeek and Anthropic’s Claude Code. Victims include Deere, SharkNinja and Cummins; NorForge’s report did not name them, and Reuters identified them by reviewing the Cairn data itself.

What the companies say: Deere said it took remediation actions and has “no indication that sensitive information was accessed.” SharkNinja said the breach hit a standalone product-support site and has been contained. Cummins said it is investigating a potential attack on a website run by a third-party vendor. Reuters could not establish how much data was taken. According to The Verge, Cairn’s creator @le1xia0 said earlier this month that the platform “was never a ‘hacking tool,’” and NorForge is working to get the malicious instances taken down.

Reactions: @choblin29 amplified the Reuters story (464 likes, 51K views); @Hesamation joked that the hackers forgot to password-protect their own agent dashboards, which is how researchers could literally watch the agents work (38 likes); David Manheim argued that generic, easily usable cybersecurity is ending — defenders either monitor proactively or any script kiddie can have an LLM attack them (personal view).

Also: Following the attack on OpenCode, OpenAI’s Tibo (@thsottiaux) said on Oct 19 that his team worked with OpenCode to get affected users who are also OpenAI API customers back in business and to refund any credits drawn during the attack. He urged defenders to request Daybreak Blue access before an incident hits (6,532 likes, 620K views).

Sources: @choblin29 · @Techmeme · @Hesamation · @davidmanheim · @thsottiaux · Reuters (via Yahoo) · The Verge · NorForge blog

3. MiniMax launches M3.1 Flash: the stealth model “Space Bunny” was an early build; $0.15/$0.60 per million tokens, weights on Oct 23

What MiniMax says: [Company claim] On Oct 19, MiniMax’s Ryan Lee posted “Dear friends, we are back” (1,108 likes), followed by an official teaser and the launch of M3.1 Flash (3,966 likes, 517K views). According to the official blog, Space Bunny was an early version of M3.1 Flash and processed more than 200 trillion tokens while anonymous. Online inference is nearly three times faster, with everyday output above 200 tokens per second and a median above 170. Pricing is flat at $0.15 input, $0.60 output and $0.015 cached input per million tokens. The official X post says it costs a third as much as M3, while the blog says overall usage cost on agentic tasks is 70% lower than the previous generation. Self-reported scores include 73.7 on DeepSWE v1.1, 62.2 on Automation-Bench, 76.2 on BankerToolBench, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, 82% on CyberGym and 91 on BabyVision with tools. It is available via the MiniMax API, M Plan and MiniMax Code; weights arrive Oct 23, and a Pro model will come later.

How it compares: According to MarkTechPost, M3.1 Flash has 428B total parameters with about 23B active and a 1M-token context; reasoning cannot be turned off and comes in five effort levels. M3 is priced at $0.30/$1.20, so Flash is half that per token. In MiniMax’s own comparison table, as compiled by CellCog, DeepSeek V4.1 Flash-1017 scores higher on DeepSWE (77.0 vs 73.7) and Terminal Bench 4.0 (42.3 vs 38.1), while M3.1 Flash leads on Automation-Bench (62.2 vs 60.5). @kimmonismus summed it up: the price matches DeepSeek’s off-peak rates and halves its peak rates, but cached input costs more, and it trails the latest DeepSeek on four of six comparable benchmarks (454 likes).

OpenRouter: OpenRouter said Space Bunny was its No. 1 model for 10 straight days and processed 56.7 trillion tokens on the platform, the most of any stealth model it has run. It left on Oct 5 and returned today as M3.1 Flash, free for the next 72 hours (547 likes). The 56.7 trillion figure covers OpenRouter alone, while the 200 trillion cited by MiniMax is its cumulative total across channels, so the two are not directly comparable.

Early tests: [Personal test] @SPAC89 ran the same task on M3.1 Flash (via M Plan) and DeepSeek V4.1 Flash-1017. Both finished in 38 minutes at about 200 tokens per second, and he preferred MiniMax’s result (624 likes).

Also: [Self-reported by the leaderboard] Vals AI says DeepSeek V4.1 (Oct 17 version) is now the top open-weight model on the Vals Index, at $0.64 per task (245 likes).

Sources: @RyanLeeMiniMax · @MiniMax_AI 1 · @MiniMax_AI 2 · @OpenRouter · @testingcatalog · @kimmonismus · @SPAC89 · @ValsAI · MiniMax blog · MarkTechPost · CellCog (benchmark table)

4. No “Black Monday” at the open: AI chip stocks rebound and the Nasdaq rises about 1%, after the Philadelphia chip index fell 11.4% last week

The open: [Media report] At 9:36 a.m. ET (21:36 Beijing time), Barron’s reported the Nasdaq Composite up 1%, the S&P 500 up 0.5% and the Dow down 0.1% (63 points). NDTV Profit’s 9:31 a.m. readings: Dow 51,964.04 (-0.07%), S&P 500 7,703.58 (+0.54%), Nasdaq 26,424.21 (+1.05%). These are figures from the first minutes of trading, not closing levels.

Premarket: Reuters said CoreWeave, which plunged more than 25% last week, rose 6.6% before the bell; Broadcom and Micron gained more than 3%, and Nvidia and Amazon more than 2%. Intel rose 4.2% after J.P. Morgan upgraded it to “neutral” from “underweight.” The iShares Semiconductor ETF climbed 3.6% premarket after falling almost 12% last week. Bloomberg said the Nasdaq 100 had just had its worst week since the tariff turmoil of April 2025.

Last week’s damage: According to Reuters, the Philadelphia semiconductor index fell 11.4% last week, its biggest weekly drop since March 2025, and had fallen for eight straight sessions as of Friday, wiping out about $2 trillion in market value. Hot inflation that raised the odds of Fed rate hikes and doubts about whether heavy AI spending will pay off both weighed on the sector.

This week: SpaceX reports on Tuesday, and Meta and Tesla on Wednesday. Morgan Stanley strategists led by Mike Wilson describe recent moves as the mid-cycle of the AI trade, with the focus shifting from enablers to adopters (Barron’s).

Also: [Third-party claim] SemiAnalysis published an analysis saying datacenter capacity will grow by about 50GW in 2027, with 60GW under construction and $1.3 trillion in lease contracts signed so far this year — and that the shortage could move to chips (412 likes).

Sources: @Investingcom · @business · @FirstSquawk · @SemiAnalysis_ · Barron’s live coverage · NDTV Profit · Reuters (via Investing.com, premarket) · Reuters (via Investing.com, chip stocks) · SemiAnalysis

5. Guarding against agents as insiders: Microsoft expands Project Perception, and Claude Code catches a cryptominer on a Mac

Microsoft: [Company claim] In an Oct 19 blog post, “Securing the agentic insider,” Microsoft argued that frontier models, “regardless of their origin, cannot be trusted,” that enterprise agents are often overprivileged, and that Guardian Agents should supervise every action before it happens. Satya Nadella posted that this puts into practice his argument that models should be treated as insider risks, with security controls sitting outside the model itself (412 likes). According to GeekWire, the new features keep a running risk score for each agent; when risk rises, the system can limit the agent to reading data, require a person to approve a high-stakes step, or shut it down. The number of agent types it can detect on employees’ computers rises from 25 to 60, including Claude Code, Codex and GitHub Copilot. The features are in private preview, with broader availability planned for Ignite on Nov 17–19.

The cryptominer: [First-person account] A Reddit user asked Claude Code to help build a Mac temperature-monitoring app, and Claude found xmrig using “about 1175% CPU, roughly 12 cores,” with the CPU peaking at 112°C. “Your Mac is running a crypto miner that I didn’t put there,” it told him. Tom’s Hardware says the miner was probably added to the startup items on Aug 29; Claude asked permission before stopping the startup item and the process and moving the files into a quarantine folder, and the user then ran a Malwarebytes scan. @kimmonismus’s post about the episode drew 7,541 likes and 419K views.

The other side: hidden instructions in email. @aiwithsally warned that instructions hidden in phishing emails can trick agents into doing things they were never supposed to do (169 likes). NEAR co-founder Illia Polosukhin (@ilblackdragon) said such attacks are part of their AttackBench and that effectively all agent harnesses are susceptible in some form, including to running arbitrary code on the agent’s machine (248 likes). These are personal views.

Sources: @satyanadella · @kimmonismus · @tomshardware · @aiwithsally · @ilblackdragon · Official Microsoft Blog · GeekWire · Tom’s Hardware

6. Developer Monday: T3 Code says 60% of its Codex users also use Claude Code; Gergely Orosz on trusting AI outside your expertise

The tool wars: Theo (@theo) said that over the past two weeks, more than 60% of T3 Code users who used Codex also used Claude Code (1,733 likes) [Company claim]. @BrendanPlayford said Codex was unusable all weekend after Friday’s usage “fixes,” burning through three banked resets and more than 60K credits across two accounts, and that he is cancelling all his subscriptions and going back to Claude (447 likes) [One-sided claim]. Allie K. Miller’s take: Claude leads on models and behavior, Codex on features and modalities (138 likes).

The review loop: The Claude Code team’s code-review approach, published last week (Claude writes about 90% of the code; merged PRs more than doubled from June to September while the share reverted fell by more than 40%), kept spreading over the past 24 hours as a flowchart. @boristane called it “the only sane code review loop” (875 likes); @VicVijayakumar said “every single org is building this loop right now,” though big companies can’t do the last step because of regulatory hurdles (566 likes); Kent C. Dodds marveled that some people at Anthropic still write 10% of the code by hand (233 likes).

“Gell-Mann amnesia”: Gergely Orosz said that in areas where he is an expert he spots the AI’s errors immediately, but outside them he tends to trust it. He asked three models about a legal question; all three confidently gave the same answer, his lawyer gave a different one, and the lawyer was right (1,178 likes). Others dubbed this the Gell-Mann amnesia effect. Vlad Mihalcea added that Opus 5.5 usually writes what he would, but every now and then makes “terrible design decisions that would fail in production” (448 likes). These are personal experiences.

Sources: @theo · @BrendanPlayford · @alliekmiller · @boristane · @VicVijayakumar · @kentcdodds · @GergelyOrosz · @vlad_mihalcea · Claude Code team code review blog

In brief

• Hossenfelder: don’t make human understanding a requirement for science: Physicist Sabine Hossenfelder argued that continued progress means accepting that the correct description of nature will eventually be too complicated for humans to understand, and that “AI is the way forward”; a critic said this has the same logic as “humans cannot understand the mind of God,” and @politicalmath said using an LLM to work through an idea gave him “empty intellectual calories” — all personal views (@skdh, 3,620 likes, 163K views · @ToKTeacher, 126 likes · @politicalmath, 1,164 likes).
• The fight over blocking agents continues: Yuchen Jin compared it to Steve Jobs initially refusing third-party iPhone apps — “You can’t stop a tech revolution”; OpenCode’s Dax Raad said his team barely uses hosted AI assistants, and because OpenCode drives local browsers it isn’t as easy to block as “all of Muse’s ip range” (@Yuchenj_UW, 240 likes · @thdxr, 172 likes).
• Nvidia releases Kimodo on Hugging Face: a controllable human-motion diffusion model that turns text and kinematic constraints into 3D human and robot motion, trained on 700 hours of motion-capture data (@HuggingPapers, 556 likes · Hugging Face model card).
• Nebius takes its AI Builder Program to Europe: [Company claim] $10 million in total compute grants, capital and community support, kicking off this week at GTC Berlin; Jensen Huang stressed that “entire industries will be built on open models” (@nebiusai, 442 likes, 253K views · @nvidia, 271 likes).
• Mollick: retire the “a computer can never be held accountable” slide: he said automated systems already make management decisions all the time (trading, logistics, customer service, insurance) and pretending otherwise “launders responsibility” (@emollick, 758 likes).
• Musk on persuasive superintelligence: “So many people are going to get one-shotted by super persuasive super intelligences and I might be one of them” — the second most-engaged AI-related post of the past 24 hours; his own remark, without elaboration (@elonmusk, 45.4K likes, 4.13M views).
• Musk keeps promoting Grok Bot: “Hire Grok @Bot today. You won’t regret it.” (@elonmusk, 7,352 likes, 2.67M views).

---

About this column and how we verify

• AI-assisted: AI Daily Brief is compiled with AI assistance; this edition was generated with the help of Grok Bot and covers 22:05 Beijing time on Oct 18, 2026 to 22:05 Beijing time on Oct 19, 2026.
• Every post checked: Each cited post was read individually to confirm it exists, that the author matches and that it was published within this edition’s time range; all cited posts were also re-read through the X API, with authors and timestamps matching. Like and view counts were read at about 22:15 Beijing time on Oct 19. US stock figures are from the first minutes of trading, not closing prices.
• X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts rest on the original posts, official pages and media reports, and where a summary conflicts with a primary source, the primary source wins.
• Items involving Grok Bot: This column is produced with Grok Bot’s help; the “Musk keeps promoting Grok Bot” item in In brief is included purely on engagement.
• What the labels mean:
  • [Personal pledge by Musk; no details announced]: a statement Musk made on social media, with no details on how or when it would be carried out.
  • [Media report]: from news reports, not confirmed point by point by the parties through their own channels.
  • [Media exclusive]: an exclusive by a single outlet, based on a researcher’s report and the outlet’s own checks.
  • [Company claim]: figures or statements published by the company itself, not independently verified.
  • [Personal test]: a test run by an individual user, with a small sample and uncontrolled conditions.
  • [Self-reported by the leaderboard]: rankings and data published by the operator of an evaluation leaderboard.
  • [Third-party claim]: estimates and judgments by a research or analysis firm.
  • [First-person account]: a description given by the person involved.
  • [One-sided claim]: one party’s account, not confirmed by the other side.