Everything that mattered in AI, one page a week.

Google released Gemini CLI on June 25, an agent harness that runs Gemini 2.5 Pro locally in the terminal with built-in Model Context Protocol support, Google Search grounding, GEMINI.md system-prompt files and bundled extensions, plus access to Veo and Imagen for media generation. It is open source under Apache 2.0, and signing in with a personal Google account gives 60 model requests per minute and 1,000 per day at no cost against 2.5 Pro's 1-million-token context window; heavier use routes through Vertex AI or AI Studio keys with usage-based billing.

The harness ships an explicit sandbox story rather than leaving it to the user: native macOS Seatbelt confinement, execution inside Docker or Podman containers, and network traffic routed through proxies for inspection. Google's open-sourcing argument is auditability, and in a week when a rival's IDE extension needed a WebSocket patch, the ability to read the code that mediates tool calls is the substantive part of the release, not the license label.

The Wall Street Journal reported on June 25 that Meta had hired Lucas Beyer, Alexander Kolesnikov and Xiaohua Zhai — the three researchers who set up OpenAI's Zurich office after leaving Google DeepMind — for Zuckerberg's superintelligence effort; an OpenAI spokesperson confirmed the departures. The following day TechCrunch reported that Trapit Bansal, at OpenAI since 2022 and a foundational contributor to the o1 reasoning model's reinforcement-learning work, had also left for Meta's reasoning team. Zuckerberg's offers to individual researchers have been reported at $100M-plus packages.

The context is a company buying capability it has not been able to ship: Meta has no publicly available reasoning model, Llama 4 landed in April to poor reviews, and its access to frontier talent was previously bought at the company level — a $14.3B investment for a 49% stake in Scale AI in June, which brought CEO Alexandr Wang with it. Google, in the same 48 hours, published open weights that run on a phone for free.

DeepMind published AlphaGenome on June 25, a sequence model that takes up to 1 megabase of DNA as input and predicts thousands of functional genomic tracks at single-base-pair resolution — gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription-factor binding, contact maps and splice-site usage. The preprint reports it matching or exceeding the strongest external models on 24 of 26 variant-effect evaluations, and reproducing the mechanism of clinically relevant variants near the TAL1 oncogene. Science and Nature both covered it the same day.

Access is the interesting part for builders. The model is free through an AlphaGenome API for non-commercial research, with client code public on GitHub, but source code and weights are promised only once a peer-reviewed version lands, and a commercial offering is in early testing. DeepMind also says the API suits analyses of thousands of predictions, not millions. It is the inverse of the week's open-weight releases: capability published, artefacts gated, commercial path provisional.

The CVE record for GHSA-9f65-56v6-gxw7 was published on June 24: Claude Code's VSCode extension, versions 0.2.116 through 1.0.23, and the JetBrains beta plugin, 0.1.1 through 0.1.8, lacked origin validation on the WebSocket their local server opens, rated CVSS 8.8 by GitHub's CNA. Because the vulnerable extensions are also used in VSCode forks, Cursor, Windsurf and VSCodium users were in scope. A page the developer merely visits can script a connection to that local server.

What an attacker gets is the agent's own context: arbitrary file reads, the list of open files, selection and diagnostics events, and limited code execution where a Jupyter notebook is open and a malicious prompt is accepted. Anthropic patched on June 13 — 1.0.24 for VSCode-family extensions, 0.1.9 for JetBrains — and auto-update covers most installs, but the disclosure only became public the following week.

The Model Context Protocol published its 2025-06-18 specification revision, which requires a protected MCP server to implement OAuth 2.0 Protected Resource Metadata (RFC 9728) and advertise its associated authorization server, and requires clients to send RFC 8707 resource indicators on token requests so access tokens are audience-bound to one server. That closes the confused-deputy path where a token minted for one MCP server could be replayed against a more privileged one in the same deployment. JSON-RPC batching was removed in this revision, and the protocol version is now enforced via an HTTP header.

The same revision added typed tool output and elicitation — a server can send an elicitation request with a JSON schema to ask the user for missing input mid-session. A companion security best-practices document leads with the confused-deputy problem and calls for per-client consent, exact redirect-URI matching and OAuth state validation. The practical upshot is that MCP servers stop being bespoke credential stores and start plugging into whatever identity provider already issues tokens for your organization.

The Wall Street Journal reported Monday that OpenAI executives have discussed filing an antitrust complaint with US regulators against Microsoft over the terms of their contract, and Reuters reported Wednesday that Microsoft is prepared to abandon the restructuring negotiations if the two sides cannot agree on critical issues including the size of Microsoft's future equity stake in OpenAI — Microsoft plans to rely on its existing commercial contract for access to OpenAI technology until 2030. The reported ask is a 33% stake in a restructured entity in exchange for forgoing rights to future profits, alongside changes to Microsoft's exclusive rights to host OpenAI models on Azure.

The structuring fight is not new — it follows Meta's December request to California's attorney general to block OpenAI's conversion to a for-profit, and Elon Musk's suit to halt it — but the escalation is: OpenAI needs Microsoft's approval to complete its transition into a public benefit corporation, and both companies are now publicly stating that talks are ongoing and amicable. Microsoft has invested about $13 billion in OpenAI since 2019.

MiniMax released MiniMax-M1 on June 16 under Apache 2.0: a hybrid-attention mixture-of-experts model with 456B total parameters, 45.9B active per token, a 1M-token context window and an 80K-token thinking budget. It reports 55.6% and 56.0% for the M1-40k and M1-80k variants on its SWE-bench validation run, below DeepSeek-R1-0528's 57.6% but ahead of other open weights, and leads open models on the TAU-bench agent tool-use set. The headline number is the training bill: three weeks of reinforcement learning on 512 H800s for $534,700 in rental cost, which Fortune reported as roughly 200x below published estimates of GPT-4o-class training spend.

API pricing is $0.40 per million input tokens up to 200K context and $1.30 per million between 200K and 1M, with output at $2.20 — MiniMax says the first tier undercuts DeepSeek-R1 while the long-context tier has no DeepSeek equivalent. It was not a lone release: Moonshot published Kimi-Dev-72B the same day under MIT, claiming 60.4% on SWE-bench Verified and state of the art among open coding models. Two permissively licensed models, one day, both aimed at agentic coding.

The Defense Department's Chief Digital and Artificial Intelligence Office awarded OpenAI a contract with a $200 million ceiling, disclosed Monday: a pilot to develop prototype frontier AI capabilities addressing critical national security challenges in both warfighting and enterprise domains, with just under $2 million obligated at award and an estimated completion date of July 2026. OpenAI used the moment to launch OpenAI for Government, an initiative that also consolidates existing work with the Air Force Research Laboratory, NASA, Los Alamos and the Treasury Department.

The contract is roughly 2% of OpenAI's claimed $10 billion annual recurring revenue, and it sits awkwardly against OpenAI's own usage policies, which bar using its services to develop or use weapons — The Verge flagged that the money is for warfighting domains while OpenAI says all use cases must comply with its guidelines. Breaking Defense noted the award names CDAO, a Pentagon office with unfilled billets and persistent consolidation rumors, as the contracting activity.