Automated AI News Brief: Gemini 3.8 Flash, Copilot Governance, and MCP

Introduction

Horizon collected the source material for this post, while Codex selected and rewrote it. Horizon is used only for data collection.

1. Gemini 3.8 Flash Arrives, with llm-gemini Support

Google released Gemini 3.8 Flash, alongside 3.8 Flash Cyber, which is available only to trusted defenders. llm-gemini 0.34 now adds gemini-3.8-flash with low, medium, and high thinking levels, and fixes async responses failing to record the resolved model version. When a new model enters a toolchain, teams should compare speed, cost, and output quality with fixed tasks before deciding to migrate based on a single demo.

Sources: Google: Gemini 3.8 Flash and 3.8 Flash Cyber, Simon Willison: llm-gemini 0.34

2. Copilot App and CLI Now Support Content-Exclusion Policies

GitHub Copilot app and Copilot CLI now respect content-exclusion policies configured by enterprise, organization, and repository administrators. Excluded files are not used as context. This makes it possible to isolate sensitive configuration, credential examples, internal documents, or other content that should not be sent to a model at the source level. Teams still need to review rule coverage regularly, since exclusions work only when paths and file classifications are correct.

Source: GitHub Changelog: Content exclusions generally available in Copilot app and CLI

3. Enterprise Settings Can Set Any Default Copilot Model for New Conversations

GitHub now lets enterprise-managed settings select a preferred Copilot model as the default for new conversations, allowing organizations to choose a model that fits their workflows. That does not mean every task should be locked to one model. A practical approach is to define simple guidance for common jobs, such as fast completions, long-horizon agent work, and sensitive code, while preserving the ability to switch when needed.

Source: GitHub Changelog: Enterprise-managed settings support any default model

4. datasette-mcp 0.2 Returns SQL Results as Arrays of Objects

datasette-mcp 0.2 is the plugin's first non-alpha release. It changes execute_sql so its rows result is an array of objects instead of an array of arrays, helping weaker models avoid losing track of which positional value maps to which column. It also now depends on mcp>=2.1.1. For agents, readable results that carry field names are often more valuable than a smaller payload, because clarity directly affects reliability after a tool call.

Sources: Simon Willison: datasette-mcp 0.2, GitHub: datasette-mcp 0.2

5. openai-python 3.7.0 Updates Usage APIs and Documentation

openai-python 3.7.0 updates the usage APIs and documentation, and fixes repeated NumPy checks for embeddings. Even a small SDK release can affect usage-data parsing and monitoring reports. Projects with custom usage ingestion or cost dashboards should validate field handling and test data before upgrading.

Source: openai-python 3.7.0

Takeaway

Today's theme is that agent toolchains are starting to resemble mature enterprise software. Alongside new models, they are making content boundaries, default selection, and tool-result readability more explicit. Those less glamorous settings often affect whether teams can use agents safely and reliably more directly than a single jump in model capability.