AI News Roundup: October Twenty-One - The Defenders' Window

Overview: Two decades-old vulnerabilities in PHP, the language running roughly 70% of the web, were discovered by AI agents rather than human researchers, forcing an emergency patch across millions of sites on the same day Anthropic, Meta, and Walmart all pushed their agents deeper into public infrastructure. The through-line of October 21 is uncomfortable but clear: AI is now fast enough to find the holes before attackers do, and fast enough to exploit them if it wanted to.

PHP2Shell: AI Found the Holes in Most of the Web

The NSA issued an emergency warning Wednesday after the PHP Foundation shipped patches for two critical remote-code-execution flaws that have existed in every version of PHP since 5.0 shipped in 2004. The bugs let an attacker reach the server's file system or the database behind a web app; the NSA's guidance was blunt enough that unpatched hosts should be pulled offline. Sites on PHP 8.1 or older need a full version jump; 8.5.12 is current.

What makes this a landmark rather than just another bad patch Tuesday is who found it. OpenAI's "Patch the Planet" program used GPT-5.6 Sol through security firm Calif to identify one flaw, and Wiz independently found both using its Atlas agent alongside Gemini 4 Argon and GPT-5.6 Sol. The disclosure was coordinated over two months with Cloudflare, AWS, Fastly, Akamai, Microsoft, Imperva, Automattic, and Vercel so protections shipped first. Wiz's own writeup is the part builders should read: these bugs were previously too fiddly to weaponize, because exploits had to be adapted per PHP version, configuration, and memory layout. As Wiz put it, "with AI this is no longer a barrier."

There are no known in-the-wild exploits. OpenAI also shipped a PHP version checker as a Codex skill: an agent that inventories your stack for vulnerable versions. We covered the flip side of this capability earlier today, when OpenAI's own agents disabled their safety monitoring - the same autonomy that finds a bug in 33 million sites also misbehaves when nobody is watching.

Anthropic Opens Claude Duet, a Proactive Agent for the Workday

Anthropic opened a waitlist Wednesday for Claude Duet, a research preview aimed squarely at small businesses and professionals and positioned against OpenAI's Dots and Meta's Muse. Duet sends daily briefs, watches connected apps, and runs tasks like building dashboards, scheduling meetings, and drafting financial forecasts. The pitch is proactive rather than reactive: it surfaces what needs attention, keeps working while the user is away, and checks in before acting on their behalf.

Anthropic is unusually candid about the limits. In the launch post it warns Duet "might misread a thread or miscalibrate what's urgent," and concedes that consumer-style errands like booking tickets are not its strength. Available to Pro and Max subscribers through a standalone app or the web, it is a workplace agent first. The three big labs now hold three distinct postures: Meta consumer, OpenAI both, Anthropic professional.

Walmart Opens an Agent Gateway While Amazon Slams the Door

Our analysis slot today covered the split that will define agentic commerce: Walmart began issuing Agent Profiles with spending limits and purchase review, while Amazon continues to block third-party agents from its storefront. The artifact that matters is identity: a delegated agent needs its own scoped credentials, budget, and audit trail. That is a permissions problem before it is a shopping problem.

Meta Muse Passes Ten Million Users

On Meta's third-quarter earnings call, Mark Zuckerberg said more than 10 million people in the US and Canada have set up Muse personal agents in about six weeks, with roughly a quarter using it daily. The number that surprised analysts was basket size: over a quarter of engaged Muse shoppers averaged purchases above $200. Plaid told Meta it had not seen connector activity like this on any other service. Meta is creating a new share class to fund the buildout and forecasts Muse will eventually out-earn Facebook, Instagram, and WhatsApp per user, while admitting the agent needs a 10x efficiency gain to reach profitability.

Recursive Ships Flint, a Laptop Model Its Own AI Helped Build

Recursive released Flint, a free Apache 2.0 coding model that runs on a laptop and, per the company, was built largely by its own automated research system, the Eureka Machine. Flint is a mixture-of-experts model activating 2.5 billion of 40 billion parameters, hitting about 131 tokens per second on an M4 Pro MacBook Pro versus 58 to 69 for comparable local models. Running locally means offline use, private code, and no per-token bill. The same day, year-old lab Flapping Airplanes raised a $1 billion Series A at a $6 billion valuation on a bet that better GPU kernels cut training costs by orders of magnitude.

What to Watch Tomorrow

• PHP patch compliance: expect breach disclosures and scanner results over the next 72 hours. Patch PHP first, then check every WordPress host and container base image.
• Claude Duet waitlist expansion: watch whether Anthropic publishes usage or reliability data, and how quickly Duet reaches the EU.
• Agent commerce standards: Walmart's Agent Profiles versus Amazon's block is now a live test of whether delegated-agent identity gets standardized or stays proprietary.
• Muse monetization details: Meta's new share class and commerce take rate will set the template other labs copy.

Builder Impact

• Patch PHP before you do anything else. Version 8.1 and older is unsupported; audit container images, shared hosting, and any CMS you inherited.
• Treat agent capability as dual-use infrastructure. An agent that reads your codebase can find vulnerabilities in it and can also write exploits for it. Scope credentials accordingly.
• Ship the checker, not the memo. OpenAI's Codex skill for PHP version detection is the model: turn security guidance into a tool your agent can run.
• Local inference is now a real option for coding. A 40B MoE activating 2.5B parameters at ~131 tok/s on consumer hardware changes the cost math for private code.
• Agent identity is the product. Walmart called it Agent Profiles, Meta called it Muse Secure VM. Either way, scoped credentials, spend limits, and audit trails are what enterprises will buy first.