Independent trackers logged no new model launches in the last 24 hours, so the day belonged to enterprise strategy rather than product. Two items still reach builders: Anthropic tripled the reach of its Glasswing security programme, and Google and Microsoft lined up an enterprise agent-protocol alliance aimed squarely at Anthropic's MCP. Two big launches, OpenAI's GPT-Live on 8 July and Meta's Muse Spark 1.1 on 9 July, landed just before this window and are noted below for context.
Project Glasswing expands to 150 organisations across 15 countries
If you run security for critical infrastructure, this is the item worth your attention today. Anthropic tripled the reach of Project Glasswing, its programme that deploys the restricted Claude Mythos cybersecurity model to find and automatically patch vulnerabilities in the codebases societies depend on, from 50 partner organisations to 150 across 15 countries. It matters because autonomous AI attack tooling is no longer hypothetical: researchers have documented JADEPUFFER as an end-to-end autonomous ransomware operation, so patching at machine speed is becoming an operational requirement rather than a conference demo.
The honest caveat is concentration risk, since 150 critical organisations now lean on one vendor's model and disclosure pipeline. This is a gated deployment programme, approved organisation by organisation, so it is not something you can switch on yourself, but it is the clearest sign yet that AI-versus-AI defence is now a live operational discipline.
Google and Microsoft line up an enterprise agent-protocol alliance against MCP
If you are deciding how your AI agents connect to enterprise data and tools, that choice just got more contested. Google, Microsoft, Salesforce, Snowflake and ServiceNow have agreed to back a shared backend protocol for AI agents, framed explicitly as a counter to Anthropic's Model Context Protocol and to OpenAI in the enterprise. Between them these five touch most of the world's enterprise data and the clouds it runs on, so a coherent, co-supported standard would hand businesses a credible alternative to building their agent stack on a rival's standard.
The complication is that the same companies, along with Anthropic and OpenAI, also sit inside the Linux Foundation's Agentic AI Foundation, so they are cooperating on open standards and knife-fighting in the market at the same time, and MCP's roughly 18-month head start is larger than the announcement implies. For anyone building agents right now, the practical takeaway is to keep your tool-connection layer abstracted rather than betting the whole stack on one protocol while this shakes out. Whoever ends up owning the agent-connection standard gets default status in enterprise deployments for years, which is why a plumbing story is one of the more consequential items of the week.
On models, nothing shipped from Google in the window. Gemini 3.5 Pro, with its trailed 2 million token context window and Deep Think reasoning, is expected around 17 July but is not yet confirmed with a model card or pricing, so it counts as anticipation rather than a release.
Today's themes
The frontier labs front-loaded their product news into the 8 and 9 July window, with GPT-Live, the GPT-5.6 family, Grok 4.5 and Muse Spark 1.1 all landing there, which left the 13th dominated by enterprise strategy and legal noise. The one development that actually reaches builders is the Google and Microsoft agent-protocol alliance, a sign the competition has moved from model quality to the plumbing that connects agents to enterprise data, where Anthropic's MCP currently leads. Security is the other live thread, with Anthropic's Glasswing expansion pointing at machine-speed patching as a new operational requirement in a world of autonomous attack tooling. Much of the week reads as positioning ahead of the expected 17 July Gemini 3.5 Pro launch, so this stretch is the quiet before a likely loud day.