Patch first, read the postmortem later. The lead in AI news today is a web-security emergency: PHP released fixes for two critical flaws that can allow an unauthenticated attacker to run code on a server. The project rates both at 9.2 under CVSS 4.0 and says there is no complete mitigation short of updating. If your organization operates PHP directly—or relies on a host that does—this is an operations task, not a bookmark-for-Friday story.

The rest of the day shows why the underlying workload keeps getting harder. Anthropic previewed an agent that works while you are away. OpenAI crossed a competitive boundary by putting its models into Google Cloud. Meta reported $34.91 billion in quarterly capital expenditure and prepared a new non-voting share class. Nvidia offered a modular route to building AI capacity faster, while U.S. communities demanded more transparency before approving it. Capability is accelerating; permission, patching and public consent are trying to catch up.

AI news today: PHP ships emergency fixes for two critical RCE flaws

On October 21, the PHP project released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix CVE-2026-109618 and CVE-2026-109619. Both can permit unauthenticated remote code execution. PHP’s notice is unusually blunt: there are no complete mitigations, and updating is the only safe choice. PHP 8.2 also reaches end of life on December 31, adding a second deadline for teams still running it.

Palo Alto Networks’ Unit 42 says it has seen proof-of-concept material and estimates, from its own telemetry, that more than 22.6 million exposed instances may be vulnerable. “Potentially” matters: an internet scan does not prove a specific server is exploitable. The practical response is still immediate—inventory versions, patch every runtime and container image, restart workers, then review logs for suspicious activity. Do not stop after updating one visible WordPress site if the same server hosts several PHP applications.

Claude Duet turns the assistant into a background worker

Anthropic opened a waitlisted research preview of Claude Duet, a standalone iOS, Android and web app that monitors connected accounts, prepares work and continues tasks on its own virtual machine. It starts with U.S. Pro and Max subscribers and shares their normal Claude usage limits. Anthropic says Duet asks before sending messages, paying, signing, deleting or posting—and forbids some actions, including securities trading and moving money between accounts.

The product is early enough that Anthropic also warns it can misread threads or misjudge urgency. That caveat is the story. A proactive agent can remove coordination work, but it can also amplify a bad priority without waiting for a prompt. Builders should watch approval logs, connector permissions and the quality of “needs your attention” judgments. For context, our analysis of how Anthropic uses AI internally shows why workflow design matters as much as the model.

OpenAI crosses into Google Cloud’s Model Garden

OpenAI and Google Cloud announced that OpenAI frontier models and Codex workflows will become available through Google Cloud’s Model Garden and Marketplace. Customers will be able to apply existing Google Cloud spending commitments to purchases; general availability is planned in the coming weeks. OpenAI also says its models are coming to Google Antigravity and Gemini Agent, while Microsoft remains its primary cloud partner.

This is distribution, not a merger of the rival labs. For enterprise buyers, it reduces procurement friction and makes model comparison easier inside an existing governance perimeter. It also turns cloud commitments into a competitive lever: the “best model” may be the one your contract can already buy. What comes next is the less glamorous evidence—exact model availability, regional limits, pricing, data-handling terms and whether Codex workflows behave consistently across clouds.

Meta’s Q3 numbers put a price on the AI buildout

Meta reported third-quarter revenue of $64.60 billion, up 26% year over year, and capital expenditure of $34.91 billion. Free cash flow was negative $2.12 billion, a non-GAAP measure, while the company narrowed its full-year capex forecast to $135–145 billion. Meta attributes incremental advertising revenue to its AI investments, but that is management’s assessment rather than a disclosed causal breakdown.

The board also approved a publicly listed, non-voting Class C share class, effective in January 2027, for capital raising, employee compensation and potential acquisitions while preserving the founder-led voting structure. The contrast with yesterday’s Gartner AI ROI gap is useful: Meta can point to a powerful ad engine, yet the infrastructure bill is arriving faster than many new AI products can be separately valued. Watch financing, free cash flow and product-level monetization—not capex alone.

Nvidia packages an AI data center for delivery

At GTC Berlin, Nvidia introduced DSX POD, a reference design for Vera Rubin systems assembled, cabled, cooled and tested inside modular containers before delivery. The company claims that a prepared site could move from roughly six months of deployment work to as little as 14 days. Partners including Dell, HPE, Lenovo, Schneider Electric and Vertiv are expected to deliver systems in early 2027.

That timeline is a vendor target, not field-proven performance. Even a perfectly integrated container still needs land, power, water, permits and network capacity. The design nevertheless attacks a real bottleneck: on-site integration. Infrastructure leaders should compare total installed cost, serviceability and local resource constraints, not simply “time to first token.” The cleverest container in the world remains, inconveniently, attached to a power grid.

Data-center politics becomes an infrastructure constraint

The Associated Press reports that Amazon, Google, Meta and Microsoft are offering host communities more transparency, fewer nondisclosure agreements or larger local investments as opposition delays projects. Amazon and Meta have each announced $1 billion in community incentives; Google says its local-program investments will reach $1 billion by the end of 2027. Critics quoted by AP remain unconvinced, and more municipalities are considering or adopting moratoriums.

This turns community trust into a schedule risk that belongs beside chips and substations. A promise is not a permit, and a community fund is not proof that electricity, water and tax costs are fairly allocated. What comes next is project-level disclosure: who pays for grid upgrades, what usage is guaranteed, which tax concessions apply and whether residents see those terms before a vote.

Watch & Learn

Video: “Introduction to Vertex AI Model Garden” by Google Cloud Tech (about six minutes). The walkthrough explains how a model catalog connects discovery, notebooks and deployment. Some model names in this 2023 tutorial are dated, but the core workflow is still useful—especially for enterprise teams evaluating what OpenAI’s arrival changes. The title, verified channel and lesson outline were confirmed; use the text link because playback embedding could not be tested reliably.

AI, Translated: remote code execution

Remote code execution (RCE) means an attacker can make a computer run commands from somewhere else—often without legitimate access. Imagine a crafted web request that persuades a vulnerable PHP worker to install malware or read configuration secrets. “Unauthenticated RCE” is especially serious because the attacker does not need an account first. You should care because the vulnerable component may sit underneath an application you trust; updating the app’s content or password does not repair the runtime executing it.

Try This Today: turn two security bulletins into one patch plan with ChatGPT

Goal: compare the vendor advisory with an independent security analysis in ten minutes. ChatGPT supports document and text uploads on Free and paid plans, subject to plan limits and workspace settings.

1. Save or paste the official PHP notice and one independent advisory into a new chat.
2. Ask for a table separating confirmed facts, vendor claims, unknowns and required actions.
3. Verify every version number against the official notice before assigning work.

Copy-ready prompt: “Compare these two security bulletins. Produce a patch checklist with affected and fixed versions, evidence of exploitation, assets to inventory, restart or rebuild steps, verification checks and unresolved questions. Cite the source for every row. Treat the vendor advisory as authoritative for versions, and label all inference.” Do not upload confidential logs or credentials; use sanitized excerpts.

One thing to remember

The week’s scarce resource is not intelligence. It is controlled execution: patched runtimes, explicit approvals, measurable economics and communities that understand what they are being asked to host.