GitHub Copilot Updates Five Systems at Once

Over two and a half months—from July to September 2026—GitHub rolled out five Copilot updates at once: a new metrics API for agent features, a prompt injection detector in CodeQL 2.26.0, GPT-6 Astra in GA, the GPT-5.6 trio (Sol, Terra, Luna), and Gemini 3.6 Flash.

Individually, they are a routine changelog list. Together, they send a signal: the question “which model should we deploy”

gives way to the question “how do we measure and secure what is already working?”

CodeQL 2.26.0 added the js/system-prompt-injection query, which detects cases where untrusted external data reaches an LLM system prompt. Support for Kotlin 2.4.0 arrived as well. For teams already embedding LLMs into their products rather than merely using Copilot as an assistant, this is the first static analyzer of its kind in mainstream tooling—previously, injection checks were performed manually or not at all.

A seemingly simple outcome: “we found a vulnerability before production”

- relies on serious engineering work: data flows between external input and the system prompt must be modeled precisely to avoid drowning in false positives.

Measurement (metrics API), protection (CodeQL), and choosing the right tool for each task (the model lineup) are the building blocks of a mature AI environment for development.

A one-off deployment of a trendy assistant falls short of this kind of environment.

Without metrics, a company cannot see where it is losing money on unused integrations.

Without injection analysis, it cannot see where it is losing control of the agent.

Without a deliberate choice of model for each task, a company overpays for capacity where it is unnecessary and cuts costs where savings come at the expense of quality. In projects where we deploy an LLM & Security Gateway and an MCP environment for a client at KT.Team, these three layers—usage metrics, injection checks, and routing requests to a model with the right cost and capability—become part of the infrastructure from day one, rather than being added after an incident or budget audit.

The business does not measure the fact that Copilot with GPT-6 is installed somewhere.

The business measures the outcome: a specific backlog feature reached production faster, without data leaking where it should not.