Copilot gets a model refresh and a security filter
GitHub has rolled out updates to both Copilot for Individuals and Copilot for Business, centered on a more capable AI model and a new real-time vulnerability filter. The changes are intended to improve the accuracy, speed, and safety of code suggestions directly in the editor.
The headline change is an upgrade to the underlying Codex model, which GitHub says yields large-scale improvements in suggestion quality while reducing latency. This follows earlier work on the model since Copilot's initial release in June 2022. At that time, roughly 27% of developers' code files were, on average, generated by Copilot; that figure now sits at 46% across all languages and jumps to 61% for Java.
To achieve these gains, GitHub made several technical adjustments beyond the model swap. The company adopted Fill-in-the-Middle (FIM), a prompt-construction paradigm that uses both the prefix and suffix of the code being edited, leaving a gap in the middle for Copilot to fill. This gives the model more context about the program's intent and how new code should fit with the rest of the file. GitHub reports that FIM consistently produces higher-quality suggestions and was engineered to avoid adding latency.
GitHub also updated the Copilot extension for VS Code to include a lightweight client-side model that filters what gets shown to the user. Using basic signals—such as whether the previous suggestion was accepted—this model suppresses unwanted suggestions. This change reduced irrelevant prompts by 4.5%, improving overall acceptance rates. A second iteration of that client-side model shipped in January 2023 brought further gains.
Blocking insecure patterns as you type
Alongside the model upgrade, GitHub introduced an AI-based vulnerability prevention system designed to stop insecure coding patterns at generation time. It targets common issues including hardcoded credentials, SQL injections, and path injections.
Instead of a traditional static analysis pass, this system uses LLMs to approximate the behavior of static analysis tools. Since Copilot runs on powerful compute resources, the check is fast enough to run in real time and can even flag vulnerable patterns in incomplete fragments of code. When such a pattern is detected, it is blocked and replaced with a safer alternative suggestion.
The filter is trained to recognize patterns that appear vulnerable—for instance, code that mimics keys or passwords in a way that presents an unsafe credential-handling pattern—rather than necessarily showing real credentials. GitHub is asking developers who encounter vulnerable suggestions to report them at [email protected], using the reports to train the LLMs to better distinguish vulnerable from non-vulnerable code.
This capability complements GitHub Advanced Security's detection and remediation features, creating a workflow where prevention happens in the editor and remediation can follow later in the pipeline. All of these changes are now available to Copilot for Individuals and Copilot for Business subscribers.
