Cursor Just Declared War on GitHub Over AI Agents

🤖 AI Agents

Cursor Just Declared War on GitHub Over AI Agents

What happened: Cursor — the AI coding tool now owned by SpaceX — launched Origin, a git-based code-hosting platform built directly into its editor. It’s an explicit alternative to GitHub, with repositories, pull requests, and stacked-PR support baked in, rolling out now in early beta to paying customers.

Why it matters: The launch landed the same day GitHub suffered a multi-hour outage across Actions, API, and Copilot. That’s not a coincidence — Cursor is betting that code hosting needs to be rebuilt for the agentic era, not bolted onto it. The pitch is “agent-native infrastructure”: source control that actively collaborates with your AI agents, auto-moving pull requests toward a mergeable state rather than passively waiting for a human to review.

What’s next: Watch two things. First, whether Cursor can convert its massive editor install base into Origin users without alienating teams that still rely on GitHub as their system of record (Origin mirrors GitHub repos but doesn’t replace pushes). Second, how GitHub responds — it’s already building its own agent workflows and an “AgentHQ” hub. This is the opening salvo in a fight over who owns the developer’s workflow when the developer is partly an AI.

Source: InfoQ

The Single-Model Era Is Ending — Companies Are Now Routing Agents to Different Models

As agents take on more complex workflows, businesses are moving beyond one flagship model and building routing systems that dispatch each task to the best-suited model. The “so what”: your agent stack is becoming an orchestration problem, not a model-selection problem — cost and capability now get tuned per-task. Business Standard

Anthropic Locks Down Agent-to-Tool Access for the Enterprise

Anthropic launched enterprise-managed authorization for Model Context Protocol (MCP) connectors, letting organizations control and secure how their agents reach external tools and data. This is the governance layer agents have been missing — expect security teams to start treating agent connections like they treat API keys. CyberPress

📰 AI News

Anthropic Poaches Google’s TPU Founder — the Silicon Arms Race Is On

Anthropic hired Amir Salek, the man who founded and led Google’s custom TPU chip program through seven generations, to spearhead its push into self-designed silicon. It’s the clearest signal yet that AI labs now see owning the hardware stack — not just renting GPUs — as the next decade’s strategic battleground, directly mirroring OpenAI’s Broadcom “Jalapeno” chip push. T客邦

All 21 Top Open-Weight Models Can Be Stripped of Their Safety Guardrails

A University of Waterloo and FAR.AI study found every one of 21 popular open-weight LLMs could be fine-tuned to bypass its safeguards — enabling mass disinformation or harmful instructions at scale. The researchers released an open-source tool called TamperBench to stress-test models, warning that procurement of AI for government services needs far more rigorous, evidence-based review. Mirage News

OpenAI Quietly Brings Back the Five-Hour Usage Cap on ChatGPT Plus

OpenAI reinstated a five-hour usage quota for ChatGPT Plus subscribers on Codex and ChatGPT Work features — a limit that had been removed after the GPT-5.6 release. It’s a reminder that “unlimited” AI access remains a moving target as providers manage inference costs, and worth factoring into any workflow that depends on heavy sustained usage. MSN