Agent-written code is now outrunning the infrastructure and the budgets behind it.

Today's headline

Agent-written code is now outrunning the infrastructure and the budgets behind it. GitHub says it handles 7.4 billion commits a month and had to add four million CPU cores to cope. Alphabet fell more than ten percent on a cash-flow miss, and Microsoft trimmed its spending forecast. Meanwhile, security teams are still cleaning up after PHP2Shell.

On the clock

Friday brings four dates. Mistral retires one of its GLM models, so move callers to the newer one. OpenAI Evals goes read-only, so export your saved evals. DeepSeek's free search and fetch promotion ends. And Mistral has promised open weights for Large 4 by the end of the month. They were not out as of this morning.

The next date is November fourth, when Moonshot promises open weights for Kimi K3.1 Pro. Line up an evaluation now, and read the license before you plan to self-host.

Security

PHP2Shell is now rooting firewalls. It is critical, it is exploited, and PHP has fixed releases. A China-linked group found about nineteen thousand firewalls, chained PHP2Shell with a Linux kernel bug, and took over twenty-five of them. That includes a network tied to the US government. Palo Alto confirms attacks on exposed management interfaces. Take management interfaces off the public internet this week, and upgrade PHP.

Cisco has patched a firewall flaw that attackers are using to crash VPN devices. It is high severity, and it needs secondary authentication on a tunnel group. If you run those firewalls, upgrade to a fixed release now.

Alpine Linux's GitLab server was probed with a critical flaw, scored ten out of ten. The surprise is that upgrading GitLab did not fix it. The fix lives in the Ruby that GitLab ships, and Alpine ran its own. If you run GitLab on your own Ruby, upgrade Ruby too.

CISA added a VMware ESXi bug to its exploited list. The vendor rated it low severity, because it only hides what an administrator does. CrowdStrike says a China-linked group used it as a zero-day. Patch by exploitation status, not by score.

Two vLLM advisories are rated high. One lets a forged shared-memory handle reach a pickle load, but only when that cache mode is on. Upgrade vLLM to version zero point thirty-two on those hosts. There is no CVE score yet, so scanners may stay silent.

Last, a researcher found about twenty-six thousand working cloud and AI keys in public Gitee repositories. Deleting a key from code does not revoke it. Rotate anything that was ever pushed.

Agents

Google DeepMind released Gemini 4 Carbon two weeks after Argon. It tops Artificial Analysis's new index at 59, though it still trails Claude Opus 5.5 on coding agents. Because of its cyber strength, it starts with selected security experts.

GitHub made Copilot Fusion generally available. It is one entry in the model picker that routes and orchestrates models for you. GitHub says its top setting scored within two points of Claude Code with Opus 5.5, at about half the cost. Those are GitHub's own figures.

Android Studio now lets more than forty outside coding agents run inside the IDE. That includes Claude and Codex. Each one gets access to the project structure, the emulator and the logs. Google also opened a platform for registering MCP servers with Search and Gemini, in US beta. Cloudflare also opened a beta web search API in AI Gateway, so search calls get the same logging and billing as model calls.

Models and pricing

GitHub Enterprise now includes Copilot for every licensed user, with no included credits. Usage is metered from the first request against your enterprise budget. Set budget alerts before you turn it on.

Nvidia released Nemotron 3.5 Transcribe, an open speech model. It handles forty languages and runs on a single GPU. Nvidia's own benchmarks put it near the top.

Artificial Analysis rebuilt its index, and sixty percent of it now uses private tests. The gap between open and closed models grew from three points to twelve. Kimi K3.1 Pro fell from fourth to ninth. Recheck any model choice you made on the old rankings.

Business and policy

Alphabet fell more than ten percent after reporting negative free cash flow of nine billion dollars. Search growth was about twelve percent, against fifteen expected. Microsoft cut its capex forecast to one hundred seventy billion dollars and said third-party GPU delays push some capacity into next year.

Microsoft also disclosed that Azure brought in thirty-three billion dollars, and that OpenAI arrangements produced ten billion in one quarter.

Anthropic's IPO filing shows less than a third of its first-half compute bill went to serving customers. Most of it went to training and internal use. Tom's Hardware estimates about eight hundred billion dollars more in compute commitments. And Oracle will offer Google's TPUs, the first time they are available outside Google.

AI in production

GitHub says September commits hit 7.4 billion, and agents now start over forty percent of pull requests. Expect heavier CI load and slower queues on shared runners.

Anthropic's Claude Code team says Claude writes about ninety percent of its code, and nobody needs a full mental model of it. Practitioners are pushing back, since no one is left to debug by understanding. One pull request went through more than fifty review rounds overnight, so the team capped them.

Amazon told engineers to stop chasing token counts, a practice it calls tokenmaxxing. Its guidance says usage metrics become flawed once they are targets. If your dashboards reward usage, expect the same gaming.

Action items

Here are the actions from today.

Take firewall, VPN and gateway management interfaces off the public internet this week, and upgrade PHP.

Upgrade Cisco ASA and FTD to a fixed release if any tunnel group uses secondary authentication.

Upgrade Ruby to a fixed release on any GitLab you run yourself, and GitLab to nineteen point three point six.

Upgrade vLLM to zero point thirty-two on hosts that use the shared-memory cache, and keep Ghostscript off image-serving hosts.

Revoke any cloud or AI key that was ever pushed to a public repository, because deleting it from the code does not help.

Before you go

Start with your firewalls. PHP2Shell is exploited, critical, and aimed at management interfaces, so closing that exposure is the one task here that should not wait for a meeting. The rest of the list can follow in the order above.