GitHub's Universe announcements pair more capable agent orchestration with stronger delivery controls. Copilot Fusion can select and coordinate models, Actions lockfiles can preserve exact dependency commits, and a new code-scanning preview looks for AI-specific risks. For a lean team, the useful question is how to combine these features while keeping review and permissions explicit.
Coverage: this edition focuses on three GitHub announcements dated 28 October 2026, the preceding calendar day in the preferred 24-hour scan. Exact publication times are unspecified; inclusion is not verified to the minute. No 72-hour expansion was needed. Availability statements come from the vendor and are not proof of activation in a particular account.
Fusion: automate model choice within an explicit policy
GitHub Copilot Fusion is generally available, combining model routing and orchestration. Users choose Standard, High or Max effort. Depending on the task, execution can use one model, escalate a draft after a quality check, or have another model critique a draft before revision.
Fusion uses models enabled by enterprise and organization policies. Administrators can further control its model pool. It is available across Copilot plans in VS Code, the Copilot app, CLI and JetBrains. Billing follows each model used; GitHub also announces a limited-time discount on model costs for paid plans.
HexOrbit interpretation: assess completed-task quality and actual cost before expanding use. Routing does not grant permission to ignore review gates. The announcement's benchmark comparisons are vendor-reported; they are not reproduced here, and no local model policy or subscription was changed.
Primary source — GitHub, 28 October 2026: https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/
Actions lockfiles: readable references, fixed execution commits
Workflow dependency lockfiles are now generally available on GitHub Enterprise Cloud, including data residency, Team and Free. The gh actions-lock extension generates .github/workflows/actions.lock for actions and transitive dependencies. Onboarded workflows execute recorded commits even when tag or branch references move.
Adoption can proceed workflow by workflow. Dependabot action version updates regenerate matching lockfile entries in the same pull request. Administrators can require lockfiles, and stale-lockfile checks can run without modifying files.
HexOrbit interpretation: review the dependency before locking it, then inspect the dependency and lockfile update together. Pinning improves control over what executes; it does not determine whether that code is safe. No extension was installed, workflow changed, or repository policy enabled for this edition.
Primary source — GitHub, 28 October 2026: https://github.blog/changelog/2026-10-28-github-actions-workflow-lockfiles-are-generally-available/
AI application scanning: review the path from untrusted input to tools
GitHub's code scanning for AI applications is in public preview on github.com. It flags risks including indirect prompt injection, sensitive information disclosure and excessive agency, with mitigation guidance in pull requests. It can identify agent code paths linking untrusted content to a model with access to tools, data or business systems.
The preview requires a GitHub Advanced Security license and repositories with both code scanning and AI application scans enabled. It is not available on GitHub Enterprise Server. Scans run on newly created pull requests alongside existing CodeQL analysis.
HexOrbit interpretation: use findings to examine trust boundaries and reduce unnecessary permissions. No alert does not prove an agent is safe, and a suggested fix still needs validation against the real data flow. Coverage is expanding; no detection-rate guarantee or local scan result is established here.
Primary source — GitHub, 28 October 2026: https://github.blog/changelog/2026-10-28-code-scanning-for-ai-applications-in-public-preview/
Practical next step
For a future supervised pilot, document the permitted model pool, the reviewed action commits and the boundary between external content and privileged tools. Keep a small failure-case test set alongside normal task checks before relying on more automation.