GitHub deprecates GPT-5.2 and GPT-5.2‑Codex in Copilot

What changed: GitHub announced on June 5, 2026 that GPT‑5.2 and GPT‑5.2‑Codex are deprecated across most Copilot experiences; it recommends migrating to GPT‑5.5 and GPT‑5.3‑Codex respectively.

Why it matters: If your coding agents, Copilot scripts, or automated PR workflows explicitly select those model IDs, they may stop receiving updates or become unsupported; model behavior, latency, and cost can differ between the deprecated and replacement models, so regressions in agent outputs or token usage are possible.

Try/watch: Immediately run a small integration test suite using the suggested alternatives (GPT‑5.5 / GPT‑5.3‑Codex), check enterprise model policies and feature flags in Copilot admin settings, and compare output for key agentic tasks (refactoring, tests, release notes) to catch regressions early.

VS Code gets enterprise‑managed plugins for Copilot/CLI (public preview)

What changed: On June 5, 2026 GitHub published a public‑preview changelog showing VS Code (and Copilot CLI) can now accept enterprise‑managed plugins that administrators configure and auto‑install via a central settings file; the feature supports plugin types that let teams distribute vetted agents, skills, and MCP connector configs across users.

Why it matters: For platform teams and security-conscious buyers this is a meaningful way to scale agent capabilities without asking every developer to install extensions manually: you can centrally push approved agent plugins (or turn off risky ones) and reduce onboarding friction for agentic workflows.

Try/watch: Pilot by packaging one internal agent/skill as a managed plugin and test auto‑deployment to a controlled cohort; use the enterprise settings.json approach to enforce MCP connector defaults (so third‑party data sources are consistent), and watch for plugin update/rollback behavior before broad rollout.

Microsoft Defender: agentic CI/CD workflows are a new high‑risk attack surface (Claude Code GitHub Action case)

What changed: Microsoft Defender published a June 5, 2026 technical post describing how an AI‑action (the Claude Code GitHub Action) could expose CI/CD secrets when untrusted repo content is processed by an agent; Anthropic addressed the specific issue in Claude Code v2.1.128 but Microsoft urges treating any agentic CI workflow that reads untrusted inputs as high risk.

Why it matters: Coding agents and developer automation are now part of build pipelines — if an agent tool or its “Read” tool can access files or environment variables, it can inadvertently leak credentials or trigger privileged actions. That risk changes how you design agent access and runtime permissions for build and deploy steps.

Try/watch: Audit all CI/CD workflows that invoke agents: remove agent access to secrets and file reads unless strictly necessary, prefer scoped tokens, pin to patched action versions (or vendor‑verified forks), and add observability (logging and alerts) around any agent tool that can perform I/O or network calls. Track vendor advisories for further patches.