Satya Nadella stood up at GitHub Universe this week and said something that should stop you for a second: more than 40 percent of pull requests opened on GitHub in September were opened by an agent, not a person typing at a keyboard. GitHub's own Octoverse report backs him up, and the trend line matters more than the single number. In March it was 17 percent. Six months later it's more than double that, and the report says those agent-authored PRs touch more files per change than the human-authored kind, 87 percent versus 79 percent, so they're not all small drive-by commits. Something is genuinely shifting about who does the typing.
What GitHub shipped alongside that data point is more interesting than the data point itself, though. Copilot Fusion went generally available on October 28, and it isn't another model. It's a router. Instead of you picking a model for a task, you point Copilot at Fusion and it decides which model, or combination of models, handles the work, at whatever effort level you set, across VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. GitHub is betting its moat isn't "we have the best model," because it doesn't, and everyone in this market knows that changes every few months anyway. The bet is that the orchestration and the workflow around it, issue to PR to review to merge, is the sticky part. If agents are about to author half of everything landing on GitHub, owning the thing that decides which model gets the job looks like a better business than being one of the models competing for it.
I'd hold off on the triumphant read of that 40 percent figure, though. A pull request being opened by an agent isn't the same claim as it being good, reviewed properly, or understood by whoever actually merges it. GitHub's number measures authorship, not judgment, and the gap between those two is exactly where this gets risky at scale.
Which brings me to the other story from this week, because it's a useful corrective to getting excited about agents writing nearly half the world's pull requests. On October 27, CISA added CVE-2026-111497 to its Known Exploited Vulnerabilities catalog: an out-of-bounds write in the Brotli decoder bundled with Node.js. The bug is almost embarrassing once you read the writeup. A malformed compressed stream can report an invalid block switch as "need more input" instead of a format error, so the streaming decoder keeps appending bytes to an internal buffer that was only ever eight bytes wide. Any app decompressing untrusted Brotli data through node:zlib, meaning any server accepting compressed request bodies or any client trusting compressed responses from somewhere it shouldn't, is exposed. Depending on memory layout you get a crash, a leak of process memory, or remote code execution. Node patched it the same week across all three lines it currently supports, 22.23.4, 24.22.1 and 26.12.1, with CISA setting a federal remediation deadline of November 10.
The detail I keep coming back to is who found it: OpenAI's own Codex Security team gets the credit on Node's advisory. Which makes the week's two headlines more connected than they look. One says agents are writing a huge and growing share of the code going into production. The other is a reminder that the code everything else depends on, a compression library nobody thinks about until it breaks, still has the kind of bug that's turned up in C decoders for decades, and that it took a dedicated security team, human or AI-assisted, actually looking for it to catch.
And Node.js 26 crossed into Active LTS on that same October 28, which is a smaller note but a real one. It's also the last release on the old six-month cadence before Node moves to shipping one major version a year. If you're still running 24.x for anything serious, it isn't going anywhere yet, but the clock on upgrading has started regardless of whether an agent or a person pushes the commit that does it.