September 12, 2026: Copilot gets tighter controls, new metrics, and a model retirement

On September 12, 2026, GitHub is starting from a clean operating baseline: its status page says all systems are operational, with no active incidents listed.

The more interesting story this week is not outage risk. It is control-plane churn around Copilot.

First, GitHub expanded what enterprise admins can actually see. As of September 11, Copilot usage metrics now include VS Code Agents window activity, with daily active users and per-user and aggregate session and message rollups. In plain English, that means admins can separate agent-window adoption from broader Copilot usage instead of looking at one blended number.

That matters because agent workflows are not just another chat box. They are a different kind of product surface, and GitHub is now instrumenting them more explicitly. For enterprise buyers, telemetry is not a vanity add-on. It is the difference between guessing at adoption and seeing whether a new workflow is actually taking hold.

Second, GitHub tightened governance around those agent workflows. On September 9, Copilot Business and Enterprise admins got centralized controls for agent operations, including whether shell commands, file reads and edits, and network domains are blocked, require human approval, or proceed without a prompt. That is a meaningful step for security-conscious buyers. It lowers the gap between “we want agents” and “we can let agents touch real code and tools.” It also suggests GitHub is treating agent permissions as an admin surface, not just a user preference.

Third, GitHub changed the model menu. On September 10, MAI-Code-1-Flash was deprecated across Copilot experiences, and GitHub says MAI-Code-1.1-Flash is the suggested replacement. That is a real product change, not just a label update. For anyone managing Copilot at scale, retirements like this are a reminder that the usable model set can change underneath deployment plans. The practical takeaway is simple: model availability is a moving target, so teams need to watch the supported menu instead of assuming last week’s defaults still apply.

GitHub is also pushing more AI security automation into APIs. On September 10, it added REST APIs for AI Scan for pull requests in public preview. The important detail is not just that the feature exists, but that orgs and repos can now enable or disable it programmatically. That makes it easier to roll out code-scanning behavior across a large estate, though GitHub frames it for Advanced Security customers. In other words, this is less about a flashy new scanner and more about making security controls easier to manage at scale.

On the DevOps side, GitHub Actions got a concrete least-privilege control. Also on September 10, cache-mode reached general availability, letting workflows choose read, write, write-only, or none access to the cache. The plain-language mechanic here is simple: shared caches can speed builds, but they can also become a supply-chain risk if too many jobs can write to them. Restricting write access reduces the chance of cache poisoning. This is the kind of change that is easy to overlook, but it matters in real CI environments because it lets teams narrow what a workflow can do without turning caching off entirely.

There is also a newer AI architecture story underneath all of this. On September 4, GitHub introduced HydraFusion as a research preview in Copilot CLI. GitHub says its selective multi-model workflows matched or exceeded an Opus 5 baseline in controlled offline tests while lowering estimated cost. The plain-English idea is straightforward: instead of one model doing every step, a router can use a cheaper model for drafting, another for critique, and a stronger model for harder tasks. That is GitHub trying to trade off quality, latency, and cost per task rather than betting on a single default model. The benchmark numbers are GitHub-controlled, so they are directional evidence, not independent validation.

The reliability backdrop is steady, but not spotless. GitHub’s August availability report says the company had five incidents in August and is still moving primaries to Azure while improving Actions capacity and cache headroom. That does not mean a live outage today. It does mean the platform’s operating story is still partly about hardening infrastructure while shipping more AI controls on top. So the week’s product news lands against a backdrop of active platform maintenance, not a fully settled operating environment.

One correction to keep in view: the Aug. 27 Copilot AI Model Providers incident was on Kimi K3, not Fable 5.

So the week’s pattern is pretty clear. GitHub is making Copilot more governable, more measurable, and more modular across models, while also tightening CI and code-scanning controls. For enterprises, that is the difference between a demo and a deployable system. It is also a sign that GitHub sees the center of gravity shifting from feature launch announcements to operational control surfaces: who can use an agent, what an agent can touch, which model gets picked, and how much of that activity shows up in admin reporting.

What to watch next: whether the supported-models docs show any further fallout from the MAI retirement, whether enterprise-managed agent permissions get broader adoption signals in usage or admin docs, whether GitHub publishes a hardening note for the Copilot billing and seat-activation thread, and whether HydraFusion stays a preview or moves toward a wider rollout.