AI Industry Daily Briefing — October 22, 2026

The Executive Read

The thread today is the bill for AI, paid in two currencies: money and risk. On the money side, Meta reported on October 21 that it spent $34.9 billion on capital projects in the third quarter and now expects $135 billion to $145 billion for the year, while Tesla’s filing shows its profit squeezed by AI and robotics spending. Blackstone’s head of AI told a TechCrunch conference that worries over AI-buildout debt are overblown, a day after CoreWeave shares were reported down more than 20% in recent days. On the risk side, patches shipped October 21 for two critical flaws in PHP, software that runs a very large share of the web. Wiz, the security firm that disclosed them, says it used three AI systems in the research, and OpenAI says another flaw was found with its model by a partner. OpenAI says researchers found the holes before attackers did and that no attacks in the wild had been reported when it published. Meanwhile Anthropic, which press reports say is preparing a stock listing, announced a plan to put a share of its equity into a public benefit fund. That plan needs stockholder approval and the details are still to come.

Top AI Headlines

AI Tools Helped Find Two Critical PHP Flaws; Patches Are Out, and the NSA Says Patch or Go Offline

What happened. On October 21 the PHP project released security updates, versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, fixing two vulnerabilities that can let an unauthenticated attacker run code on a server. PHP is the programming language behind many websites, including WordPress. The two flaws are CVE-2026-109618 and CVE-2026-109619 (CVE is the public numbering system for security bugs). The PHP project rates both 9.2 out of 10 on the CVSS severity scale. The first involves PHP running out of memory at the wrong moment and leaving half-built data that it later trusts. The second involves a fatal error that leaves stale memory a later request can reuse. PHP says the highest risk is on servers running the mod_php Apache module, where an attacker’s request can trigger the attack before the site’s own code runs. PHP 8.1 and older no longer receive upstream updates, though minimal backports exist for 7.4, 8.0 and 8.1 for people who build from source. Wiz Research says it found the first flaw on August 19, confirmed working attacks against Drupal and WordPress the next two days, and reported to PHP on August 24. It says it used its own Wiz Atlas agent, Google’s Gemini 4 Argon and OpenAI’s GPT-5.6 Sol, and that Gemini 4 Argon helped with discovery, exploit development and patch checking. It combined the bugs into a demonstration exploit it calls php2shell and says it worked against current WordPress, Drupal and Joomla container images. Wiz does not say attacks have been seen in the wild. According to TechCrunch, which links OpenAI’s disclosure, a security firm called Calif found one of the flaws with GPT-5.6 Sol under OpenAI’s “Patch the Planet” program, and OpenAI says no attacks in the wild had been reported when it published. TechCrunch, citing a post it links, also reports that the NSA advised updating immediately or taking affected sites offline. I could not open OpenAI’s or the NSA’s pages myself.

Why it matters. Two lines of the PHP Foundation’s post stand out. It says Anthropic’s Claude Mythos scanned PHP earlier this year through the Alpha-Omega program, which funds open-source security, and that OpenAI’s Daybreak Red tool is used to check reports and exploits. It also says AI has raised the volume of reports: 15 CVEs published since July 1, and more than 100 reports that were invalid, duplicates or could not be reproduced. People still make every decision and write every patch, the Foundation says. Wiz’s warning is the other side of the same coin: models make it cheaper to adapt an exploit to different server settings, and once a patch is public, attackers can read it too. The race is now to update the long tail of servers.

Business implication. Anyone who runs or hosts PHP sites should confirm their version today and, if they cannot update, apply the temporary settings in PHP’s advisory, which PHP says do not make a system safe. Companies that depend on PHP 8.1 or older face a harder choice, since upstream no longer patches it.

Sources: PHP.net security advisory and releases (October 21); PHP Foundation (October 21); Wiz Research, “php2shell”; TechCrunch (October 21), which links OpenAI’s disclosure and the NSA advisory. The OpenAI and NSA pages returned access errors for me; what they say is as TechCrunch relays it.

Meta Reports $64.6 Billion in Revenue and Narrows 2026 Capital Spending to $135–145 Billion

What happened. Meta’s third-quarter release, dated October 21, shows revenue of $64.60 billion, up 26% from $51.24 billion a year earlier. Net income was $9.86 billion, or $3.83 a share. That includes a $9.89 billion charge the release ties to an agreement with 52 U.S. state attorneys general; excluding it, Meta says earnings per share would have been $7.40. Capital expenditures, including principal payments on finance leases, were $34.91 billion in the quarter. For the full year Meta now guides to $135–145 billion of capital spending, narrowed from $130–145 billion, and $175–179 billion of total expenses, raised to include the legal charge. It guides fourth-quarter revenue to $73–76 billion. Reality Labs, its headset and glasses unit, is expected to lose $18–19 billion in 2026 and about $15 billion in 2027.

On the call, according to TechCrunch, Mark Zuckerberg said Muse, Meta’s AI agent app, has more than 10 million users and has reached “product-market fit.” He said Meta forecasts Muse will eventually earn more per user than its family of apps, that current revenue per user is “several hundreds of dollars annually,” and that Meta expects to improve Muse’s compute cost efficiency by more than 10 times within a year. TechCrunch reports about 25% of people who set up Muse use it daily and that Muse will stay free for most users, with Meta taking a share of commerce.

Why it matters. Meta’s spending now sits inside a single, checkable range, and the company is telling investors the agent product is the reason. The usage numbers and the 10x cost claim are Meta’s own statements on a call, not audited figures, and the legal charge is not about AI.

Business implication. Merchants and developers building for agent shopping have a new large channel to consider, with Meta saying it takes a percentage of commerce. Suppliers to Meta’s data centers should read the narrowed range as a floor of $135 billion, according to the company’s own guidance.

Sources: Meta Investor Relations, Third Quarter 2026 Results (October 21); TechCrunch on the Muse call remarks (October 21). Call quotes are as relayed by TechCrunch; I did not read a transcript.

Tesla’s Operating Income Falls 56% as AI and Robot Spending Rises

What happened. Tesla’s filing for the third quarter shows revenue of $28.0 billion, roughly flat on the year. GAAP operating income was $718 million and net income attributable to common stockholders was $787 million. Research and development rose to $2.8 billion from $2.4 billion in the second quarter, and management cited its “SI” work (Tesla’s term for its AI training effort) and other projects. Capital spending was $4.7 billion, down from $5.8 billion in the second quarter. Free cash flow was $1.5 billion. Tesla says it plans to reach 400 megawatts of SI training compute by year-end, with two clusters in production at more than 105 and 120 megawatts. Of the humanoid robot Optimus, it says production has begun on high-volume lines in Fremont, with Texas under construction. It says Cybercab, its two-seat robotaxi, has begun joining the Austin fleet, and that cumulative unsupervised Robotaxi miles are nearly two million. Electrek, which analyses Tesla’s letter, calculates that Tesla’s earlier “more than $25 billion” capital spending target for 2026 would need roughly $12 billion in the fourth quarter, and notes the letter does not update that target. Electrek puts the 56% drop in operating income down to AI and Optimus costs growing faster than gross profit. One site’s transcript of the call suggests a changed guide of about $20 billion; I could not verify that against Tesla’s own materials, so I am not relying on it.

Why it matters. Meta and Tesla both show the same pattern from different directions: AI costs arrive first, and the revenue they are meant to produce is stated in the future tense. Tesla’s company-reported figures for compute and robot production are targets and status claims, not audited output.

Business implication. Readers evaluating Tesla’s autonomy and robot plans should separate what is installed (clusters in production, a fleet in Austin) from what is planned (400 megawatts, Texas lines, the 2026 capital spending target).

Sources: Tesla Form 8-K, Exhibit 99.1, Q3 2026 results (SEC); Electrek (October 21). Electrek is an independent analysis and its 2026 capital spending arithmetic is its own.

Anthropic Plans to Commit 5% of Its Shares to a Public Benefit Fund

What happened. In a post dated October 18, Anthropic said that, subject to stockholder approval, it is committing 5% of its shares to the Anthropic Public Benefit Fund, a new wholly owned limited liability company. It said the board can add 2% more for every 50% increase in company value above a baseline it did not give in the post. It said none of the equity will be used for lobbying or political giving. It counts the pledges of its co-founders (80% of their equity, it says) and a donation-match program in which about half of employees take part, and says roughly 20% of its shares will be dedicated to the public good. It says more details on structure, governance and valuation milestones will come in the coming months, and that it is publishing its first public benefit report. On October 21, it also said physicist Christopher Stubbs, a Harvard professor and former dean of science, has joined its Long-Term Benefit Trust, a body independent of management and investors that can appoint a majority of the board.

Separately, Bloomberg reported on October 19, and TechCrunch relayed on October 20, that Eric Schmidt holds about 1% of Anthropic, with Dario Amodei at about 2% according to Reuters. Neither Schmidt nor Anthropic responded to TechCrunch. That reporting also says an Anthropic listing is widely expected before year-end.

Why it matters. The company has not confirmed publicly that it will list, though the October 20 edition covered press accounts of a draft filing. A pledge of equity before any listing fixes how much stock is set aside before public investors buy any. The post gives no valuation baseline, so the size of the potential 2% additions is unknowable from it. These are Anthropic’s own descriptions; the fund does not yet exist and needs approval.

Business implication. For customers and partners, the post is a governance signal rather than a product change. It adds one more structure, the trust and the fund, to read alongside Anthropic’s eventual public filing.

Sources: Anthropic, “Investing in the public benefit” (October 18); Anthropic, Christopher Stubbs joins the Long-Term Benefit Trust (October 21); TechCrunch on Schmidt’s stake (October 20), citing Bloomberg (October 19). The stake reporting is Bloomberg’s and Reuters’s; I did not open their pages.

Blackstone’s AI Head Says Debt Fears Are Overblown, as CoreWeave and Firmus Wobble

What happened. Speaking at TechCrunch Disrupt on October 21, Jas Khaira, who leads Blackstone’s AI hub BXN1, said, “We’re the largest owner of compute outside of the hyperscalers today,” and argued of AI infrastructure debt that “This doesn’t feel like something where you’ve got 2008 banks that are shaky credits.” He said most data-center debt is backed by hyperscaler customers, and put recent jitters down to higher interest rates, inflation signals, a flood of AI debt offerings and wider credit spreads. TechCrunch reports CoreWeave’s debt has grown to about $50 billion and its shares fell over 20% in recent days, and that Australian GPU provider Firmus, which Blackstone and Coatue backed with a $10 billion credit facility, scrapped a planned IPO at a roughly $30 billion target valuation. Khaira also said returns on AI adoption are real but “uneven.” He described a $1.5 billion venture with Hellman & Friedman and Anthropic, called Ode with Anthropic, that supplies engineers to help companies deploy AI.

Why it matters. Khaira is an interested party: Blackstone has lent to and invested in these companies. The share moves and the cancelled IPO are the counter-evidence TechCrunch cites.

Business implication. Buyers of GPU capacity from newer cloud providers should ask who the provider’s own customers are and how the debt is structured, which is the question Khaira’s argument turns on.

Sources: TechCrunch (October 21). This item rests on one outlet’s report of on-stage remarks; I found no Blackstone transcript.

Model and Product Updates

Recursive releases Flint, a free laptop coding model. Recursive, the AI lab led by former Salesforce chief scientist Richard Socher, published Flint on Hugging Face under the Apache 2.0 licence, which allows commercial use. It is a mixture-of-experts model, meaning only part of it runs on each token: 40 billion parameters in total, 2.5 billion active. Recursive says it reaches more than 125 tokens a second on an M4 Pro MacBook Pro and leads open mixture-of-experts models under 50 billion parameters on agentic coding tests including SWE-Bench Verified. Those are the company’s own results. The company says an automated research system, the Eureka Machine, helped design nearly every part of the build, including the architecture, a 4-bit quantization recipe (a way to shrink a model) and Apple-silicon kernels, and that a small team finished in under four months. TechCrunch reports Recursive acknowledges Flint needs more guidance than large paid assistants.

LiteLLM’s 1.107.0 release was malicious. Security firm Aikido reports that version 1.107.0 of LiteLLM, a popular open-source gateway for calling AI models, was published on PyPI, and as a Docker image, with a credential-stealing program and a remote-access tool. Aikido says it was found on October 21, that PyPI has quarantined the release and the images were removed. Anyone who installed or pulled it should treat cloud keys, SSH keys, CI secrets and stored AI API keys as stolen and rotate them. LiteLLM sits where many companies keep every model key in one place, which is why this one matters.

ChatGPT adds Locked Chats. According to TechCrunch, the feature, which began rolling out October 20, protects chosen conversations behind a passcode or Face ID or Touch ID. Locked chats have their own memory. They do not add end-to-end encryption, change retention or alter the “improve the model” setting, and resetting the PIN permanently deletes them. OpenAI’s own page returned an error for me.

Walmart opens an “Agent Gateway.” TechCrunch reports Walmart announced on October 21 a server using the Model Context Protocol (a standard way for AI agents to call outside tools) letting approved agents search inventory and buy at Walmart and Sam’s Club. In November it will add Agent Profiles, so each agent’s purchases appear separately, with optional per-purchase approval or spending limits. Agents that act beyond the shopper’s intent are blocked, TechCrunch reports. TechCrunch contrasts this with Amazon, which has begun blocking agents from buying for shoppers. I could not find Walmart’s own announcement.

Sources: Recursive, “Introducing Recursive Flint”; TechCrunch on Flint (October 21); Aikido Security on LiteLLM (updated October 22); TechCrunch on Locked Chats (October 20); TechCrunch on Walmart (October 21).

Regulation and Policy Watch

EU: AI Act deadlines have already moved, and December 2 is the next date. I could not open the Commission’s own page on the AI Omnibus, so this rests on secondary summaries that reference it. They agree the amending law, Regulation (EU) 2026/1744, entered into force on July 27. They say it pushes the main high-risk obligations to December 2, 2027, and to August 2, 2028 for AI built into regulated products, and that two new prohibitions, on non-consensual intimate imagery and child sexual abuse material, apply from December 2, 2026. They add that the Commission opened a consultation on October 12 on how the AI Office runs fining proceedings, open until November 9. Check the Official Journal text before acting on any date.

U.S.: the DeepSeek order stands. Yesterday’s edition covered the Department of Homeland Security order to remove DeepSeek products from civilian agency systems. Nothing new has been published that I could find on a compliance deadline.

Sources: European Commission, “AI Omnibus enters into force”, which I could not open and which is cited here for the entry-into-force date as summaries report it; Federal Register, DHS notice, 91 FR 66346.

Emerging Startup Radar

Blacksmith raised a $66 million Series C at a $1 billion valuation. Sequoia led, per TechCrunch. Blacksmith sells faster infrastructure for continuous integration, the automated build-and-test step that code, including AI-written code, passes through before release. TechCrunch reports annual recurring revenue grew 17 times, that it has run more than 90 million jobs, and that Mercury and Granola are customers. The round came two months after a $45 million Series B at $550 million, and total funding is $125 million. Growth figures are the company’s, as relayed.

Recursive, maker of Flint, emerged from stealth in May with $650 million at a $4.65 billion valuation, TechCrunch reports, with GV, Nvidia and AMD Ventures among backers.

Sources: TechCrunch on Blacksmith (October 21); TechCrunch on Recursive (October 21). Both rounds rest on TechCrunch; I found no company announcement for Blacksmith I could open.

AI Infrastructure and Market Signals

Meta and Tesla together. Meta’s own range implies at least $135 billion of capital spending this year, nearly all tied to AI, and its $34.9 billion in one quarter is larger than Tesla’s $4.7 billion by a factor of about seven. Tesla plans 400 megawatts of SI training compute by year-end. Both are company figures.

Security of the AI supply chain. The LiteLLM compromise above is the second item this week, with the PHP flaws, showing that the software around models is as exposed as the models. TechCrunch’s roundup of the day also cites a compromised Indian certificate authority, eMudhra, whose keys were reportedly exposed. Node.js, in its October 21 security releases, lists eMudhra’s intermediate certificate among its dependency updates and patches four high-severity bugs, including a Brotli decoder flaw that permits remote code execution. Node’s post does not say AI tools found any of them.

Sources: Meta Investor Relations; Tesla 8-K Exhibit 99.1; Node.js, October 2026 security releases; TechCrunch security roundup (October 21); TechCrunch on Blackstone (October 21).

Public Investment Watchlist

Informational only, not investment advice.

• Meta (META). Reported third-quarter results October 21: revenue $64.60 billion, capital spending guide $135–145 billion, fourth-quarter revenue guide $73–76 billion. I have not verified the share-price reaction.
• Tesla (TSLA). Reported October 21: revenue $28.0 billion, operating income $718 million. The 2026 capital spending target is not updated in its letter.
• SpaceX (SPCX). Reported October 20, covered in yesterday’s edition. I did not find a verified account of the share-price reaction.
• CoreWeave. Reported by TechCrunch to be down more than 20% in recent days on debt concerns.
• Intel (INTC). Reports October 29 after the close, per its investor relations page.
• Federal Reserve. The FOMC meets October 27–28.

Sources: Meta Investor Relations; Tesla Form 8-K, Exhibit 99.1; Intel Investor Relations; Federal Reserve, FOMC calendar.

Watchlist

• PHP patching. Whether exploitation in the wild is reported now that patches and the Wiz demonstration are public.
• LiteLLM. The project’s own incident write-up and how many installs pulled 1.107.0.
• Anthropic’s fund. Stockholder approval, the valuation baseline for the extra 2%, and any public filing.
• Meta Muse. Whether the claimed 10x cost improvement within a year shows up in the next quarter’s expense line.
• Tesla. Whether fourth-quarter capital spending approaches the “more than $25 billion” target.
• Mistral Large 4 weights. Mistral said they would follow by the end of October; one outlet gives October 27, which I could not verify.
• CoreWeave and Firmus. Further debt-market moves among AI infrastructure borrowers.