# The AI vs AI Battle for the Internet
*The AI Daily Brief — Thursday, 2026-10-22 · https://aidailybrief.ai/e/2026-10-22*

**Cybersecurity's new era is AI versus AI — and whoever coordinates faster wins.**

PHP2Shell was found by a frontier model, disclosed with a defender's window, leaked ten hours early, and exploited millions of times across Cloudflare's network within hours. Add the Railway breach and the OpenCode attack and the shape of the new normal is clear: swarms of cheap, relentless agents on offense versus governed fleets on defense. The old era of manageable exploits and long-held norms is ending; the new one is a coordination race fought at inhuman speeds — and long-term optimism about more secure software doesn't deny how painful the transition is going to be.

---

## By the numbers
- **70%** — Websites and web apps potentially vulnerable to PHP2Shell
- **Millions** — Exploit attempts across Cloudflare's network after disclosure
- **10 hrs** — How early knowledge of the exploit leaked ahead of disclosure
- **1.5%** — OpenCode users breached in last weekend's API-key attack
- **40%** — Higher order values from Walmart's Sparky shoppers vs non-agentic
- **25%** — Active Meta Muse users already shopping with the agent
- **$200** — Average value of a Muse agentic purchase
- **$135–145B** — Meta's data-center spend this year, low end revised up
- **$700B** — Equity Meta could still raise without Zuckerberg losing majority control

## Headlines

### Anthropic joins the personal agent race with Claude Duet `[01:00]`
Duet answers Grokbot, Meta Muse, and OpenAI's Dots with a professional and small-business twist: a proactive morning brief across your connected apps, interruptions only when something needs attention, and tasks from scheduling to forecasts to prototyping web apps. It ships as a waitlisted research preview — and for much of the community, the real differentiator was that it's built on Opus 5.5.
*For: Product*
Link: https://aidailybrief.ai/e/2026-10-22#claude-duet-joins-the-agent-race

### The agent race is now about owning the relationship, not the model `[03:00]`
Pretty much every major AI company now has a personal agent except Google — Meta's Muse, Anthropic's Duet, OpenAI's Dots, SpaceX's Bots. The competition is shifting from who has the smartest model to who owns the persistent relationship with the user, arguably because peak intelligence for most of the world's workflows was reached early this year or late last. Meanwhile nothing about the category is settled: not form factor, not interaction modes, not whether we'll have one agent or many.
*For: Product, Exec, Marketing*
Link: https://aidailybrief.ai/e/2026-10-22#agent-race-persistent-relationship

### Duet may be part one of a pre-IPO one-two punch `[04:00]`
If NLW were a betting man: release your answer to all the personal agents floating around, then give it a new model that makes everyone feel hopelessly behind. In other words, maybe Duet is just prelude to Fable 5.5 heading into Anthropic's IPO.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-10-22#duet-prelude-to-fable-55

### Walmart opens the doors to third-party shopping agents `[05:00]`
While Amazon, JPMorgan, and Bank of America say no, Walmart will let customers link personal agents with configurable spending limits starting next month — surprising given its proprietary Sparky agent already drives order values 40% higher than non-agentic shoppers. Its terms could become a template for safer agentic commerce: no actions beyond shopper intent, no modifying payment methods, approval by default, and returns protection even when a compliant agent makes a mistake.
*For: Sales, Product, Legal*
Link: https://aidailybrief.ai/e/2026-10-22#walmart-welcomes-personal-agents

### A quarter of Muse users are already shopping with it `[06:00]`
Meta's earnings call revealed that 25% of active Muse users shop through the agent, with average purchases of $200 — numbers that surprised even e-commerce veterans. Zuckerberg said Plaid told Meta it hadn't seen activity like its connector is seeing on any other service. If the early adopters of agentic shopping are this valuable, Walmart's welcome mat makes a lot more sense.
*For: Marketing, Sales, Finance*
Link: https://aidailybrief.ai/e/2026-10-22#muse-shopping-surprises-veterans

### Meta funds the build-out with equity — minus the voting rights `[07:00]`
Alongside Q3 earnings that nudged this year's data-center budget to $135–145 billion, Meta announced a new class of non-voting stock trading from January. It lets the company sell equity without diluting Zuckerberg's voting control — what CFO Susan Lee called expanding funding tools for 'the unprecedented momentum we're seeing across every part of our business driven by AI.'
*For: Finance, Exec*
Link: https://aidailybrief.ai/e/2026-10-22#meta-nonvoting-stock

### Is the bond market tapped out on AI debt? `[08:00]`
Some read Meta's equity pivot as a sign the AI debt market is exhausted; others note Meta's $900 billion in liabilities — including $350 billion in non-cancellable data center contracts — against only about $83 billion in issued long-term bonds. The Occam's razor read: Moelis says Meta needs several hundred billion in any scenario, so it's tapping every pool of capital. Shareholders apparently agree — the stock was up overnight despite the dilution.
*For: Finance*
Link: https://aidailybrief.ai/e/2026-10-22#why-equity-not-debt

## Main episode

### Project Glasswing warned us this was coming `[13:00]`
When Anthropic announced Glasswing in April, it said AI models could now surpass all but the most skilled humans at finding and exploiting software vulnerabilities — its Mythos preview had already found thousands of high-severity flaws in every major operating system and browser. The warning then: such capabilities would soon proliferate beyond actors committed to deploying them safely, with potentially severe fallout for economies and national security.
*For: Eng*
Link: https://aidailybrief.ai/e/2026-10-22#glasswing-set-the-stage

### PHP2Shell is the defender's worst nightmare `[14:00]`
The vulnerability — discovered two months ago by GPT 5.6 Sol via a researcher at Caliph and independently by Wiz — could affect as much as 70% of websites and web apps, allowing attackers to take control of backend servers and run code without logging in, even on the latest versions. PHP's Volker Dusch called it the most severe security issue the language has faced in over a decade, comparing it to 2014's Heartbleed.
*For: Eng, Ops, Exec*
Link: https://aidailybrief.ai/e/2026-10-22#php2shell-defenders-nightmare

### The defender's window closed 10 hours early `[16:00]`
OpenAI spent two months coordinating with PHP's maintainers, Wiz, and Cloudflare so that frontier AI would find the flaw before attackers did. Technically true — but knowledge of the exploit spread roughly ten hours ahead of disclosure, before anyone could patch. The NSA recommended organizations take systems offline until patched, and the FBI's Cyber Division asked for intrusion reports.
*For: Eng, Ops, Legal*
Link: https://aidailybrief.ai/e/2026-10-22#defenders-window-slammed-shut

### Even the reflexive skeptics say this one is real `[17:00]`
NLW's default is skepticism toward sky-is-falling cyber stories — but the people who know this space are rattled. One startup reported 100% of its cookies exposed, its source code stolen, its site defaced, and legal docs leaked; a former intelligence officer called it 'the calm before what is likely to be the biggest destruction event to the internet in 20 years,' with 70% of sites possibly already compromised without anyone knowing.
*For: Exec, Eng*
Link: https://aidailybrief.ai/e/2026-10-22#skeptics-say-real-deal

### Hundreds of thousands within two hours, and millions now. `[18:00]`
*— Matthew Prince, Cloudflare CEO*
Cloudflare saw thousands of exploit attempts in the ten hours before PHP2Shell was announced, tens of thousands in the 45 minutes after, and millions since — the clearest hard data on how fast AI-driven attacks now scale.
*For: Eng, Ops*
Link: https://aidailybrief.ai/e/2026-10-22#hundreds-of-thousands-within-two-hours-millions-now

### Years of vague X-risk warnings may have distracted from this `[18:00]`
Some rationalists are claiming vindication, but NLW finds that framing a little disingenuous — cyber chaos isn't close to the main thing that crowd warned about, and one could argue the focus on vague existential scenarios distracted from much more current issues like these. The best counterpoint: being concerned about one thing doesn't mean you can't be concerned about another.
Link: https://aidailybrief.ai/e/2026-10-22#xrisk-focus-was-a-distraction

### Everyone now lives and dies on patch velocity `[19:00]`
Ethan Mollick reports companies struggling with the rate of patches required as AI defenses find issues ahead of inevitable AI attacks — and consumers are next. Meanwhile much of the world is blasé: one big-tech employee, whose entire infrastructure runs on PHP, responded to the warning with 'Thanks, I'll try to patch it in a few days as I'm out of office right now.'
*For: Ops, Eng, Exec*
Link: https://aidailybrief.ai/e/2026-10-22#everyone-lives-on-patch-velocity-now

### Why they couldn't just ship the patch first `[20:00]`
Critics asked why two months of coordination produced a 'in 24 hours you're all screwed' rollout. The answer from those who know: the patch itself makes the exploit obvious, so releasing it is effectively releasing the vulnerability to a storm of script kiddies scanning everything they find. That's why the NSA told organizations to take sites down ahead of disclosure and get ready to update the moment it landed.
*For: Eng, Legal*
Link: https://aidailybrief.ai/e/2026-10-22#why-the-patch-couldnt-ship-first

### PHP2Shell may turn 15 years of old sites into lost media `[21:00]`
4chan has gone read-only and Something Awful shut its forums as sites brace for the patch; Pinboard, Choice of Games, Damn Interesting, and the Explain xkcd wiki are offline too. The lament: there's 15 years' worth of old PHP sites that may simply never come back.
Link: https://aidailybrief.ai/e/2026-10-22#the-old-internet-goes-dark

### Railway gets hit — and it's part of a pattern `[22:00]`
The AI hosting service disclosed that an unauthorized third party may have accessed encrypted customer environment variables — API keys, database and registry credentials — and decrypted some of them, with keys actively being drained. Railway believes it's part of a series of attacks targeting platform companies to compromise AI inference providers, mirroring last weekend's OpenCode exploit that breached 1.5% of users and ran up API bills.
*For: Eng, Ops*
Link: https://aidailybrief.ai/e/2026-10-22#railway-is-the-next-domino

### The only reason we're surviving is because we have agents going twenty-four seven. `[23:00]`
*— Dax, OpenCode*
OpenCode's post-exploit advice to the Railway team sums up the new normal: the attackers aren't going to stop, they'll get more sophisticated, they'll target every surface and every employee, and long-forgotten projects will be found and used against you. 'Really it's a collective problem we're all facing together.'
*For: Eng, Ops, Exec*
Link: https://aidailybrief.ai/e/2026-10-22#only-surviving-because-agents-run-247

### Where the nuclear analogy breaks: you can stop an AI with an AI `[23:00]`
One entrepreneur rotated every key across a large number of projects in minutes — an absolute nightmare not long ago — and called AI's defensive value underrated. You can't stop a nuclear weapon with another nuclear weapon, but you can stop an AI attack with an AI. The pessimistic counter-scenario, from a former OpenAI staffer: 'Feels like March 1st, 2020 for cyber. Internet lockdowns in a couple of months?'
*For: Eng, Ops*
Link: https://aidailybrief.ai/e/2026-10-22#fight-ai-with-ai-2

### Cyber may burn 100x more tokens than anyone is imagining `[24:00]`
Box's Aaron Levie argues we now have a tireless workforce that can hunt down, discover, patch, or attack any system ever created — and we'll use far more tokens on the attack-defend loop than currently imagined, scanning every log and every line of code. His jobs corollary: cybersecurity becomes one of the biggest enterprise investment areas, and even as the work goes autonomous, more talent will be needed to point agents, review findings, and respond to incidents.
*For: Exec, Eng, Finance*
Link: https://aidailybrief.ai/e/2026-10-22#cyber-is-the-100x-token-loop

### Fleets versus swarms: whoever coordinates faster wins `[26:00]`
Enclave's Tal Hoffman frames the future of cybersecurity as governed agent fleets on defense versus cheap, relentless agent swarms on offense. Others see structural consequences: markets may not be pricing in how fast banks and healthcare will have to merge just to acquire enough scale to defend against AI-driven vulnerabilities — while the 'fix your shit' camp says the only way out is through.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-10-22#fleets-vs-swarms

### Brockman floats government 'authorized defenders' that patch the internet `[26:00]`
In 'Towards Securing the Internet,' OpenAI's president says models can now find vulnerabilities with internet-scale impact and new defensive solutions will be needed — potentially including authorized defenders such as government scanning the internet for vulnerable hosts and, in extreme cases, patching those systems remotely. A remarkable proposal to sit with.
*For: Legal, Exec, Eng*
Link: https://aidailybrief.ai/e/2026-10-22#brockman-floats-authorized-defenders

### Long-term optimism doesn't deny the pain of the transition `[27:00]`
The most optimistic take — Martin Casado's — is that AI ultimately makes software far more capable, secure, and reliable, 'but before that, things are going to get much, much worse.' NLW agrees: he's an incredible long-term optimist across AI, but that doesn't deny the potential chaos, pain, and agony of the short and medium term. It's true for jobs, and it's certainly true for cybersecurity. And boy, are we in it now.
*For: Exec*
Link: https://aidailybrief.ai/e/2026-10-22#optimism-doesnt-deny-the-chaos

*Today's sponsors: KPMG, Section, CData, Blitzy — offers at https://aidailybrief.ai/sponsors*

---
Transcript: https://aidailybrief.ai/e/2026-10-22/transcript.md
Listen: https://pod.link/1680633614 · Ad-free: https://patreon.com/aidailybrief
© 2026 The AI Daily Brief — Until next time, peace ✌