Microsoft has released the Agent Governance Toolkit, an open-source project that adds policy enforcement, identity, reliability, and compliance controls around autonomous AI agents.

GitHub now lets organization admins and security managers open Copilot from Code Security and secret risk assessments, turning static findings into guided explanations and next steps.

GitHub Copilot CLI now supports bring-your-own-key model routing, local models, and an offline mode. That gives coding teams more control over cost, privacy, and where terminal agents run.

NIST has started work on a critical-infrastructure AI risk profile aimed at operators, vendors, and regulators. That could shape how AI agents enter power, industrial, and transport systems.

General Reasoning’s new KellyBench puts AI agents through a long football betting season instead of a short task list. Every frontier model tested lost money, which makes it a sharp reality check for long-running agents.

Sentence Transformers v5.4 now handles multimodal embeddings and rerankers, which means AI teams can compare text, images, audio, and video with one familiar API and less custom search glue.

Anthropic says Claude Cowork is now generally available on all paid plans, but the bigger story is the admin layer around it. New controls could decide whether larger companies roll it out widely.

Perplexity and Plaid are expanding their finance integration so users can connect checking, savings, loans, and investments for AI-generated budgets, spending views, and net worth tracking.

Meta is transferring top software engineers into a new Applied AI Engineering group, according to Reuters. The move shows how seriously the company is treating AI tooling as an internal priority.