Model routing, restricted access and the security buying gap

Copilot Fusion changes model routing and billing, Gemini 4 Carbon starts a restricted rollout, and a CISO survey examines AI security buying plans.

3 stories · Yesterday in AI

TL;DR

The short version
  1. Copilot Fusion is generally available with model-policy controls and per-model billing; automatic routing makes permitted models and accepted-task cost an administrator’s decision.
  2. Gemini 4 Carbon begins with selected cyber defenders, while broader access remains forthcoming; Google’s cyber benchmark scores were evaluated without guardrails.
  3. a16z’s 100-CISO survey reports rising budgets alongside limited AI security production use; buying intentions and useful operating capacity require separate evidence.

Enterprise Ops

1 story

Copilot Fusion puts model routing inside an administrator’s policy

GitHub’s generally available routing system combines model orchestration with administrator controls and per-model billing, creating a concrete choice about permitted models and cost.

As a leader: Decide which models Fusion may use and compare the cost of accepted tasks before setting an organization-wide default.

As a practitioner: Keep the selected effort, observable model calls, billed usage and review result with each comparison task.

GitHub ↗Vendor documentation

Vendors

1 story

Gemini 4 Carbon starts with restricted access for cyber defenders

Google’s new model pairs claimed cybersecurity gains with a phased defender rollout; announced pricing and unguarded benchmark scores do not establish broad availability or local performance.

As a leader: Confirm access and the offered safeguards before budgeting a Carbon evaluation against the organization’s own security workflow.

As a practitioner: Preserve the available model configuration, task permissions, patch checks and observed failures with the pilot result.

Google DeepMind ↗Vendor documentation

Security

1 story

CISO survey separates security buying plans from production use

An a16z survey of 100 large-company security leaders separates planned spending from reported AI security deployments, with sampling and investor perspective limiting broader conclusions.

As a leader: Require each proposed AI security purchase to name an operating gap, an accountable owner and evidence of completed work.

As a practitioner: Track investigation, approval, repair and verification effort separately when evaluating an AI security workflow.

a16z ↗Research