AI helped find two PHP flaws reaching 70% of the web, Meta put 2026 capex at $135B-$145B, and 3 more AI moves
Two critical PHP flaws that frontier models helped turn into working exploits reached 70% of the web, Meta funded its AI build-out with a new class of voteless shares, DeepSeek closed about $15 billion, Anthropic opened a proactive agent app, and Recursive shipped a coding model its own research system largely designed.
This issue covers October 21 at 12:00 UTC through October 22 at 12:00 UTC.
Five developments cleared the bar in that window.
Two critical flaws in the PHP engine were patched. Frontier models had turned a decade-old theoretical bug into working exploits.
Meta posted a record quarter, and a plan to fund AI with shares that carry no vote.
DeepSeek closed about $15 billion, one of China's largest rounds.
Anthropic opened a waitlist for a proactive agent app.
Recursive shipped a laptop-sized coding model that its own research system largely designed.
In brief
- Security: The PHP project patched two critical engine flaws, CVE-2026-109618 and CVE-2026-109619, rated 9.2. Every PHP installation that serves web requests is exposed, and Wiz puts PHP on 70% of the world's websites. Wiz found the flaws with frontier models, and says one researcher did in a week what a world-class team once needed months for. 12
- Earnings: Meta's third-quarter revenue rose 26% to $64.60 billion, while operating income fell 37% to $12.99 billion on a $9.89 billion legal charge. Capital spending reached $34.91 billion in the quarter and free cash flow was negative $2.12 billion. The board approved a non-voting Class C stock, effective next January, to finance AI. 3
- Funding: DeepSeek raised about 100 billion yuan, roughly $15 billion, in one of China's largest-ever rounds. Tencent and the battery maker CATL were among the biggest investors. The company raised its target twice on the way, from 50 billion yuan to 80 billion before the final tally. 4
- Launch: Anthropic opened a waitlist research preview of Claude Duet, a standalone proactive agent app for iOS, Android and the web. It sends a morning brief, watches connected apps and runs tasks in the background. Anything irreversible waits for your approval. 5
- Research: Recursive released Flint, a 40-billion-parameter mixture-of-experts coding model under Apache 2.0. Only 2.5 billion parameters are active per token, which is how it runs on a 48GB laptop at more than 125 tokens a second. Its automated research system chose the architecture, the quantisation and the kernels. 6
Five developments
1. AI helped turn a decade-old PHP bug into two 9.2-severity flaws
The PHP project released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 on October 21 at 14:00 UTC. 2
Those versions fix two critical vulnerabilities that allow unauthenticated code execution. 2
Both sit in memory management and error handling, the parts of the engine nearly every request touches. 2
CVSS severity is 9.2. 1
CVE-2026-109618 is a use-after-free in out-of-memory handling. 2
A request dying at the wrong moment leaves partially initialised data behind, which PHP later treats as valid. 2
CVE-2026-109619 is the mirror image. 2
Internal state survives after its memory is freed, and the next request on the same worker can use it. 2
Wiz demonstrated unauthenticated code execution against current WordPress, Drupal and Joomla container images. 1
Separately, it read passwords, tokens and API keys out of PHP worker memory without needing code execution at all. 1
The reach is the reason this leads. 1
Wiz puts PHP in the path of 70% of websites, and every version back to 5.4 carries at least one bug. 1
Affected deployments include PHP-FPM, php-cgi, FrankenPHP and LiteSpeed. 7
Apache with mod_php is the most exposed, because the flaw can fire before any application code runs. 7
Frontier models are why a fix exists now. 2
The code had been wrong for over a decade, and the exploit had been theorised since 2024 without a working path. 2
Wiz credits Gemini 4 Argon and GPT-5.6 Sol, run through its Atlas system. 1
The models found the vulnerability, developed the exploit, built detection and validated the patch. 1
Wiz's own framing is the number to sit with. 1
"A single researcher completed in a week what previously would have taken a world-class team months of work." 1
Its conclusion is blunter. 1
Complexity of this kind may have lost most of its protective value. 1
The discovery history matters for anyone deciding what to patch first. 2
PHP's first external audit in years, run by Quarkslab for the Foundation in 2024, surfaced the first signs. 2
Fixing it then stalled. 2
Every option broke backward compatibility, added a denial-of-service route, or forced changes across many extensions. 2
What broke the deadlock in 2026 was a cluster of external reports arriving within weeks. 2
Security firm Calif began auditing PHP on July 30 under OpenAI's Patch the Planet programme. 2
On August 11, researcher Ryan A at Calif reported unconditional remote code execution from the same root cause. 2
OpenAI reported the identical root cause the same day, with more vectors on September 18. 2
Wiz reported on August 24. 2
The disclosure broke before release. 2
PHP had targeted October 22 at 14:00 UTC, and the sequence below is why it shipped a day early. 2
| Time (UTC) | What happened |
|---|---|
| Oct 17 | An Amazon honeypot records an exploitation attempt, four days before disclosure |
| Oct 20, ~15:00 | The two CVE IDs circulate with a wrong release time and a wrong 10.0 score |
| Oct 20, 19:30 | PHP brings its pre-announcement forward |
| Oct 20, ~20:00 | Calif mistakenly publishes a proof-of-concept video |
| Oct 21, 07:41 | Chainguard publishes built images containing the fix |
| Oct 21, 14:00 | Coordinated release; patched versions ship |
The October 17 entry is the odd one. 7
Amazon Threat Intelligence's honeypot network recorded a PHP2Shell exploitation attempt while the bugs were still embargoed. 7
The address was tied to a commercial anonymisation service. 7
How that attacker learned the exploit before anyone was told is still unknown. 7
Chainguard's early publication drew a pointed line from the Foundation. 2
Any attacker could study the fix in binary form while every other PHP user still waited. 2
The ecosystem response was fast. 7
The UK's National Cyber Security Centre told organisations to consider taking offline anything they cannot fix immediately. 7
Cloudflare deployed firewall rules for the known exploit vectors to all customers, including free tiers. 7
Appliances are the second wave. 7
Palo Alto Networks confirmed all PAN-OS and Panorama versions are affected and has not shipped fixes. 7
Fortinet is checking 12 products and Cisco 31. 7
The Foundation's account of the delay is unusually candid. 2
"The delay was a structural failure, caused by both the security team's decision process and its lack of resources." 2
It also describes the new load. 2
The security team has published 15 CVEs since July 1 and worked through more than 100 invalid or duplicate reports. 2
AI helps sort them by root cause, the Foundation says, but a person makes every call and writes every patch. 2
What to watch: whether a working exploit appears against an appliance with no fix, and whether the Foundation's funding appeal moves its review capacity.
2. Meta's record quarter came with a $9.89 billion charge and a class of voteless shares
Meta reported third-quarter revenue of $64.60 billion, up 26% from $51.24 billion a year earlier. 3
Operating income fell 37% to $12.99 billion, and the operating margin halved to 20%. 3
Costs and expenses rose 68% to $51.61 billion. 3
Most of that jump is one line. 3
Third-quarter costs include a $9.89 billion charge tied to the August agreement with 52 state attorneys general. 3
That settlement covers teen mental health and social media, and is worth up to about $18.59 billion. 3
Strip the charge out and the quarter looks different. 3
Operating income would be $22.88 billion, up from $20.54 billion a year earlier. 3
Diluted earnings per share would be $7.40 rather than $3.83. 3
The cash line is where the AI build-out shows up. 3
Capital expenditures, including principal payments on finance leases, were $34.91 billion in the quarter. 3
That exceeded the $32.79 billion the business generated from operations. 3
Free cash flow was therefore negative $2.12 billion. 3
Guidance keeps the spending going. 3
Meta narrowed full-year 2026 capital expenditure to $135 billion to $145 billion, from $130 billion to $145 billion. 3
It raised its full-year expense outlook to $175 billion to $179 billion to absorb the legal charge. 3
The structural answer to funding that is the second announcement, and it changes who owns the upside. 3
The board approved a new class of publicly listed, non-voting Class C common stock. 3
It takes effect January 19, 2027. 3
Every Class A and Class B holder gets one Class C share through a one-time stock dividend. 3
After payment, Meta expects its Class A price to reflect an effective two-for-one split. 3
Every existing holder keeps their current voting power. 3
The new class exists to issue stock for capital raising, employee compensation and acquisitions. 3
The filings make the trade explicit. 8
Zuckerberg had signalled he could leave California if a proposed one-time 5% wealth tax passed. 8
That could have exposed him to a multibillion-dollar liability on his own shares. 8
A Founder Alignment Agreement has him staying in Meta's headquarters jurisdiction for eight years. 8
The AI products are starting to show usage, at least on Meta's own numbers. 8
Zuckerberg said more than 10 million people in the US and Canada set up Muse personal agents in the past month and a half. 8
Roughly a quarter of those who set it up use it on any given day. 8
More than a quarter of users shop through it, at an average purchase size above $200. 8
Reality Labs is still the drag. 8
The unit posted $443 million of sales against $4.13 billion of losses. 8
Meta expects 2026 losses there of $18 billion to $19 billion, narrowing to about $15 billion next year. 3
Fourth-quarter revenue guidance of $73 billion to $76 billion beat the $73.73 billion analysts expected. 38
The stock rose about 2.5% after hours. 8
What to watch: whether the Class C dividend lands without index-fund objections, and whether Muse's $200 average purchase survives its first cohort.
3. DeepSeek closed about $15 billion, after raising its own target twice
DeepSeek raised about 100 billion yuan, or roughly $15 billion, in its latest round. 4
The figures come from people familiar with the matter, not a filing. 4
The process moved, which is the part worth noticing. 4
The Hangzhou company lifted its target at least twice while talking to investors. 4
It went from 50 billion yuan to 80 billion before settling at 100 billion. 4
Tencent, which operates WeChat, and the battery maker CATL were among the biggest investors. 4
The scale needs a baseline to mean anything. 4
Anthropic and OpenAI have each raised more than $100 billion, so $15 billion narrows a gap that stays wide. 4
The context behind that gap is 2025. 4
DeepSeek showed it could compete with Silicon Valley's leaders at a fraction of the cost. 4
That wiped about $1 trillion off global markets. 4
It has spent the year since releasing high-performance, low-cost models. 4
Those pressure domestic rivals and win customers abroad. 4
The money is for data centres and further open-source models. 4
It sets up an onshore listing that Bloomberg has reported is in preparation. 4
Read this one with its provenance in view. 4
DeepSeek, Tencent and CATL did not respond to requests for comment. 4
What to watch: whether the round's final size appears in a prospectus, and whether Tencent and CATL are strategic or financial holders.
4. Anthropic's new app watches your work and asks before it acts
Anthropic opened a research preview of Claude Duet on October 21. 5
It is available by waitlist to Pro and Max subscribers in the US. 5
The app is standalone, on iOS, Android and the web, and separate from the main Claude app. 5
It is built around Claude Opus 5.5. 5
The design choice is inversion. 5
Instead of waiting for a prompt, it sends a morning brief across your connected apps. 5
During the day it messages you only when something needs attention. 5
It has its own virtual machine and browser, so it keeps working when your laptop is off. 5
The task range runs from scheduling to finance. 5
Anthropic lists dashboards across dozens of tools, financial and sales forecasts, prototyping and shipping web apps. 5
It connects through the 6,000-plus connectors in Claude Marketplace. 5
Anthropic names Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva and Shopify. 5
Payments route through Stripe's Link wallet, with a per-purchase approval and a virtual card. 5
A virtual card stands in for the real one. 5
The control architecture is the interesting part. 5
Before it sends a message, pays, signs or deletes anything, it shows you what it is about to do and waits. 5
Two things are refused even with your approval: trading securities, and moving money between accounts. 5
Credentials follow Anthropic's Managed Agents design, so the app never directly sees passwords or tokens. 5
A new classifier and safety prompt guard the strictest tier of information. 5
Credit card numbers and government identification are refused outright. 5
Health and personal finance data pass only on explicit instruction. 5
Anthropic says red teams and an external firm penetration-tested the app, and it is covered by its bug bounty. 5
The limits are stated by the company up front, which is worth crediting. 5
It "might misread a thread or miscalibrate what's urgent." 5
Personal errands such as booking tickets "aren't quite as strong," and it cannot yet show you its browser. 5
Background checks draw on the same plan limits as the rest of Claude. 5
The competitive picture is now three products. 9
Meta's Muse went consumer and reached the top of Apple's App Store. 9
OpenAI's dots straddle consumer and enterprise, while Anthropic is aiming squarely at the workplace. 9
What to watch: whether the approval gate holds for users who would rather not approve forty things a day.
5. Recursive's laptop coding model was mostly designed by its own research system
Recursive released Flint on October 21, a mixture-of-experts coding model. 6
It has 40 billion total parameters and 2.5 billion active per token. 6
The weights are on Hugging Face under Apache 2.0. 6
The company says a small team of people and agents built it from scratch in under four months. 6
The active-parameter count is the whole trick. 10
Four-bit weights take about 23GB. 10
That puts it on a 48GB Mac or a 48GB Nvidia GPU, at more than 125 tokens a second. 10
On Recursive's numbers it scores 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. 10
The similarly sized Qwen3.6 scores 73.4% and 44.9% on the same two. 10
One benchmark goes the other way, and the model card's own chart shows it. 10
Dense Qwen 3.8 scores 73.0% on Terminal-Bench 2.1, well ahead of Flint's 54.9%. 10
All benchmark figures here are the company's own. 10
One needs a caveat: SWE-bench Pro v2 is Recursive-measured, so treat it apart from public figures. 10
What makes the release more than a model launch is who did the engineering. 6
Recursive's Eureka Machine proposes ideas, implements them, tests them and picks what to try next. 6
It touched nearly every layer, from data and optimisers to training recipes, infrastructure and inference. 6
Three contributions are specific enough to check. 6
Starting from a 35-billion-parameter, 3-billion-active baseline, it searched expert configurations and widths. 6
It landed on 40 billion total and 2.5 billion active. 6
Second, quantisation: it put the output layer into 4 bits, against common practice. 6
It recovered the accuracy by distilling from the 16-bit model on a self-discovered coreset. 6
That nearly halved training steps. 6
Third, kernels: it wrote specialised kernels for low-bit projections, mixture-of-experts operations, attention and prompt processing. 6
On M4 and M5 chips it produced a Metal megakernel that fuses the entire decode pass. 6
The stated goal is recursive self-improvement. 6
The company is explicit that this is the first turn of a loop, with the loop still ahead. 6
The release supports a narrower claim than the company's framing. 6
A human team still set the target benchmarks, chose which proposals to keep, and shipped the result. 6
What remains is real and checkable. 6
A search loop cut months off a model build, at a size that runs on hardware a reader already owns. 6
What to watch: an evaluation by someone other than Recursive, and whether the next model's stack needs fewer human choices.
The read-through
The qualifier is the story.
Start with the severity score. 2
The CVE identifiers circulated on October 20 attached to a claimed 10.0; the real figure is 9.2. 2
The number that sizes that story is older than the score. 1
It is the version floor, PHP 5.4, and a flaw that sits below the application layer. 1
Meta's quarter has the same shape, twice. 3
Earnings per share of $3.83 reads as a 265% increase; excluding one charge the same line is $7.40. 3
Operating income down 37% reads as deterioration; excluding the charge it rose, from $20.54 billion to $22.88 billion. 3
Both readings sit in the same release, and the operating business grew while the balance sheet took a legal hit. 3
DeepSeek's number is the softest of the five, and soft in a way that tells you something. 4
A round reported at 100 billion yuan had a target that moved from 50 to 80 to 100 while it was being raised. 4
Demand that outruns its own target twice is a stronger signal than the closing size. 4
The figure to trust least is the one in the headline. 4
Claude Duet's word is "proactive," and the design is mostly about limiting it. 5
Every irreversible action returns to the user, and two categories never return at all. 5
An agent that asks first is slower than one that acts freely. 5
That is why the approval threshold is the number to watch. 5
Flint carries the largest number in the issue and it is the least relevant one. 6
Forty billion parameters is what makes the file big. 6
Two and a half billion active per token is what makes it run on a laptop. 6
The connection across all five is mine.
Taken together they suggest a sector where the reported figure has drifted away from the operative one. 110
Two of the five make that point about AI directly. 1
Wiz says frontier models collapsed the expertise needed to exploit a decade-old bug. 1
Recursive says an automated loop collapsed the expertise needed to design a model. 6
The PHP case is the stronger evidence, because the patch exists and the CVE record can be checked. 1
The Flint case rests on a self-reported chart and a vendor's account of its own process. 10
References
- 1
- 2
- 3Meta — Meta Reports Third Quarter 2026 Results
investor.atmeta.com
- 4
- 5Anthropic — Introducing Claude Duet
claude.com
- 6Recursive — Introducing Flint
recursive.com
- 7
- 8
- 9
- 10
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.