DEV Community

Hector Flores
Hector Flores

Posted on Originally published at htek.dev

GitHub Weekly: pull_request_target Gets Blocked, Plugin-Only Copilot, and a Model Purge

Two weeks. That's how long you have before a trigger half the open-source ecosystem relies on stops working by default. GitHub Actions will block pull_request_target in public repositories on December 2, and if you haven't looked at your policy insights, this is your reminder.

That's the headline of a week that was really about governance: who can run what, where customizations come from, and which models still exist.

What shipped

Area Change Date
Actions pull_request_target blocked by default in public repos Enforced Dec 2
Actions ARM32 hosted runner support ends Dec 8
Copilot strictPluginOnlyCustomization managed setting Nov 16
Copilot Three models deprecated Nov 16
Copilot app GitHub Enterprise Server support GA Nov 13
Security Bulk AI Scan for repositories (public preview) Nov 16
Security Dependabot dismissal requests last a year Nov 16

pull_request_target: the default flips

pull_request_target runs in the base repository's context with access to secrets. Run fork code in that context and you've built a Pwn Request. GitHub's answer, introduced with workflow execution protections going GA in September, is a default policy that blocks the event in public repos. It has been in evaluate mode; the enforcement date is December 2. Private and internal repos are unaffected, and so is GitHub Enterprise Server.

Brownouts ran on November 9 and 16 (16:00–18:00 UTC) to flush out dependencies. If your labeler or triage bot failed during those windows, you've found your list.

Your options are straightforward:

  • Switch to pull_request if you don't need secrets. The docs recommend this as the first question.
  • Explicitly allow it in an Actions event policy, using workflow file targeting to scope it to specific workflows.
  • Do nothing and let the block stand.
# Safe: triggers on fork PRs without secrets or a privileged token
on:
  pull_request:
    types: [opened, synchronize]
Enter fullscreen mode Exit fullscreen mode

My opinion: leaving it blocked is the right call for most repos. The legitimate uses, labeling and commenting, rarely need to be anywhere near fork code, and the ones that do are exactly the ones worth auditing. This is the same shift-left logic I laid out in agentic DevOps and the next evolution of shift-left: make the safe path the default and make the risky path an explicit, reviewable decision.

It matters more now that agents open and triage PRs. If you run agentic workflows, you're already giving automation untrusted input; a secure default on the trigger layer is one less thing to get wrong.

Also on the Actions front: hosted runners end ARM32 job execution on Windows and Linux by December 8, following the node20 end of support, since Node 24 doesn't support ARM32. If you have embedded-style pipelines, check now.

Copilot: plugins become the control plane

The more interesting enterprise change is strictPluginOnlyCustomization. Admins can require that customizations come from plugins or managed settings instead of standalone user or project configuration. It covers five categories: skills, custom agents, hooks, MCP servers, and executable Copilot extensions.

You can apply it to all five or pick and choose, for example locking down MCP servers and extensions while leaving skills, agents, and hooks open. Pair it with strictKnownMarketplaces to restrict which marketplaces are eligible. GitHub is explicit that plugin-only isn't the same as enterprise-managed plugins only, and it isn't a marketplace allowlist on its own.

Why I like this: until now, the choice for platform teams was often "block MCP entirely" or "let everyone wire up whatever they want." A shared plugin library with an enforced distribution path is the middle ground. It's also a reminder that custom agents are now supply chain. If you're writing them, my top 5 mistakes creating custom Copilot agents applies double when the agent is distributed org-wide.

The settings are available on GitHub Enterprise Cloud for Copilot Business and Enterprise seats.

The model purge

As of November 16, Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra are deprecated across most Copilot experiences. Suggested replacements are Claude Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna. Opus 4.8 remains for individual subscribers on annual plans, and Opus 5 survives only as a runtime refusal fallback. A further batch is scheduled for December 14.

If you've hard-coded a model name in a prompt file, agent definition, or integration, grep for it today. Admins who turned off the global default must enable the replacements through model policies; otherwise nothing shows up in the picker.

Smaller Copilot items

  • The Copilot app now supports GitHub Enterprise Server for repos, issues, and PRs. You still need a GitHub.com or GHE.com account, and cloud sessions aren't available for GHES.
  • Gemini 4 Hydrogen is GA, aimed at fast agentic and tool-use work.
  • The weekly release notes add Claude Fable 5.2 and Gemini 4 Carbon (Pro+, Max, Business, Enterprise). Fable 5.2 requires data retention by default, and Business/Enterprise admins must switch on its policy, which is off by default. Read that one before you enable it.
  • VS Code 1.146 adds an experimental Inbox mode that surfaces agent sessions needing attention, and the Local harness retires in February 2027. The Copilot harness is already the default for new sessions.

Security: AI Scan scales, Dependabot gets patient

AI Scan for repositories can now be started on up to 20 repos at once, from the org-level setting or the code scanning alerts page. It covers languages CodeQL doesn't, finds extra issues in ones it does, and lands in the same backlog as CodeQL alerts. It's public preview, and no Copilot license is required. A REST API toggle for org and repo levels arrived days earlier, so rollout can be scripted.

Smaller but sensible: Dependabot delegated dismissal requests now stay open for a year instead of a week. One week was too short for any reviewer with a real calendar. Secret scanning also now detects Cloudflare service tokens.

What it means

Read the week as one theme: GitHub is turning implicit trust into explicit policy. Triggers need an allow rule. Customizations need a sanctioned source. Models have lifecycles you must track. Security scanning is something you flip on across an org, not repo by repo.

For anyone feeding agents context, this is also the discipline behind context engineering: know exactly what goes in, and where it came from.

Bottom line

Do three things this week: audit pull_request_target workflows in Insights before December 2, grep your repos for the deprecated model names, and if you administer Copilot, decide which of the five customization categories deserve plugin-only. The direction is clear: GitHub's defaults are getting stricter as agents get more autonomous, and the teams that treat governance as code now will have the easy migration later.

Top comments (0)