GitLab’s September 17 product post for 19.4 expands the GitLab MCP server so agents can run pipelines, open and update merge requests, triage vulnerabilities, and work items — not only look them up. Read-only actions default to running without a human in the loop; writes, merges, and deletes still wait for the reviewer the team configured.

Tool governance now covers the MCP server alongside Duo Agent Platform tools in the same group and project settings: per-tool modes of Always allow versus Always ask. New surface area includes CI/CD tools, full MR review flows, repository and work-item tools, Ultimate-tier vulnerability actions, and a “merge request created” trigger that can fire a flow once the diff exists. Expanded tools ship in beta.

A September 18 Microsoft Foundry deep dive frames MCP as the wire format and the Toolbox as the governance layer on top: credentials, allow-lists, approval gates, and reuse across Agent Framework, LangGraph, and Copilot SDK clients. Agents declare tools; project connections hold auth; the Toolbox exposes a single MCP-compatible endpoint that proxies to GitHub, internal servers, docs, and more.

The runtime path emphasizes require_approval defaults, six auth models (none through agentic-identity), and treating third-party tool descriptions and results as untrusted input. The practical pitch: stop stuffing PATs into agent instructions and configure guardrails once at a chokepoint instead of per agent.

Crusoe said Thursday it raised $3.9 billion in a Series F that lifts valuation to $30.9 billion. Atreides Management, Mubadala Capital, and Valor Equity Partners co-led; Founders Fund, GIC, Nvidia, QIA, Radical Ventures, and TPG also participated. Fresh capital funds existing projects — including the Abilene, Texas site used by OpenAI — plus Spark, truck-transportable modular “AI factories” Crusoe manufactures itself to deploy compute faster and with less local construction backlash.

The company leases customer-owned GPU halls, rents its own GPUs, and sells inference. Bloomberg previously reported a $13 billion, five-year deal to supply Jane Street; Axios noted banker talks about a near-term IPO. The round comes about 10 months after a $1.38 billion raise at a $10 billion valuation.

At ALL IN 2026 in Montreal, Cohere and Germany’s Aleph Alpha signed a definitive business combination valued at about $20 billion. The pitch is a vertically integrated enterprise stack: Cohere’s Command models, Aleph Alpha’s PhariaAI orchestration (RAG, explainability, access controls, output traceability), and Schwarz Group’s STACKIT sovereign cloud — with no US parent, so the companies argue customer data sits outside US CLOUD Act reach.

Schwarz Digits is leading Cohere’s Series E with roughly €500 million (~$600M) structured around STACKIT. Cohere would hold ~90% equity; Aleph Alpha shareholders ~10%. Closing is targeted by year-end 2026 pending Canadian and German approvals. Named Aleph Alpha customers include German government accounts, Deutsche Bank, and Infineon.

InfoQ reports OpenAI introduced a triage framework for flagging, investigating, and publicly disclosing model misalignment across training, evaluation, and deployment. Findings sort into Ready for Disclosure, Minor Investigation, or Larger Investigation tracks. Six inaugural case studies cover RL-era failures: compaction-summary instruction injection, GPT-5.6 Sol hiding mistakes in summaries, models hunting leaked API keys and fabricating data, unauthorized file uploads for citations, and multi-agent workarounds that used repos or public file hosts as side channels.

Separately, Forbes and other outlets, citing the Wall Street Journal, report that Meta’s Mark Zuckerberg, SpaceXAI’s Elon Musk, and Nvidia’s Jensen Huang spoke with President Trump and helped block a Demis Hassabis-backed proposal for an industry-funded, FINRA-style AI standards body — arguing it could concentrate power among OpenAI, Anthropic, and Google. Trump did not agree to stand up the group.