EN PT-BR
← Voltar ao feed
low

GitHub releases Copilot Fusion AI model router across Copilot plans

1 fonte(s) · 1 atualização(ões) · publicado 29/10/2026 09:12 · coletado há 3 dias

GitHub launched Copilot Fusion, a runtime model router and orchestrator that automatically selects AI models and execution patterns (Single, Cascade, Critique) for coding tasks. It is generally available across all Copilot plans in Visual Studio Code, Copilot CLI, the GitHub Copilot app, and JetBrains IDEs, with local model routing (MAI Code 1.1 Flash, Qwen3.8 Flash Next) in public preview. Fusion respects administrator policies, subscription entitlements, data-residency and FedRAMP-related controls, excluding ineligible models from the picker. From a security standpoint, the article notes one operational caveat: discarding an AI draft does not reverse file edits the draft already made, so developers must review workspace changes before committing. The piece references previously reported Copilot prompt-injection issues, coding-agent jailbreaks and Agentjacking attacks, but stresses these are separate issues with no demonstrated vulnerability in Fusion itself.

Detalhes técnicos

  • Routing uses capability signals (reasoning, coding, debugging, tool use) plus system health, availability and caching boundaries.
  • Tiers: Fusion Standard, High, Max, plus a non-selectable Fusion Fast mode for VS Code inline chat.
  • No independent context window; capacity reflects the smallest limit among participating models.

Remediação

  • Review workspace diffs before committing when a draft is discarded.
  • Enforce organizational model policies and monitor AI credit consumption on High/Max tiers.

Linha do tempo

  1. gbhackers.com ↗ · pub. 29/10/2026 09:12 · coletado 29/10 11:01 primeiro registro

    GitHub releases Copilot Fusion AI model router across Copilot plans

    ler a análise desta fonte

    GitHub launched Copilot Fusion, a runtime model router and orchestrator that automatically selects AI models and execution patterns (Single, Cascade, Critique) for coding tasks. It is generally available across all Copilot plans in Visual Studio Code, Copilot CLI, the GitHub Copilot app, and JetBrains IDEs, with local model routing (MAI Code 1.1 Flash, Qwen3.8 Flash Next) in public preview. Fusion respects administrator policies, subscription entitlements, data-residency and FedRAMP-related controls, excluding ineligible models from the picker. From a security standpoint, the article notes one operational caveat: discarding an AI draft does not reverse file edits the draft already made, so developers must review workspace changes before committing. The piece references previously reported Copilot prompt-injection issues, coding-agent jailbreaks and Agentjacking attacks, but stresses these are separate issues with no demonstrated vulnerability in Fusion itself.

    Technical Details

    • Routing uses capability signals (reasoning, coding, debugging, tool use) plus system health, availability and caching boundaries.
    • Tiers: Fusion Standard, High, Max, plus a non-selectable Fusion Fast mode for VS Code inline chat.
    • No independent context window; capacity reflects the smallest limit among participating models.

    Remediation

    • Review workspace diffs before committing when a draft is discarded.
    • Enforce organizational model policies and monitor AI credit consumption on High/Max tiers.