Anthropic launched Claude Cowork on mobile and web in beta for Max subscribers, with tasks now running on Anthropic's cloud rather than a local desktop.

• Background execution means work continues after a user closes their laptop or app, with iOS and Android notifications for approvals and results.
• Internal usage data shows over 90% of Cowork activity is non-coding business operations, prompting Anthropic to extend doubled usage limits through August 5.

The US Commerce Department cleared GPT-5.6 for full public release after a weeks-long national security review, with Sam Altman confirming a Thursday launch for the Sol variant.

• Three variants ship: Sol, Terra, and Luna, with early tester Ethan Mollick describing Sol as faster and more collaborative than Anthropic's Fable, which he called more autonomous but self-directed.
• National security concerns centered on GPT-5.6's ability to identify exploitable code weaknesses, leading the government to first limit access to trusted US-based partners before approving broad release.

Elon Musk confirmed SpaceXAI will release Grok 4.5 publicly on July 9, describing it as an Opus-class model that is faster and cheaper than current frontier offerings.

• Built on the 1.5T V9 foundation model, Grok 4.5 received additional training on Cursor coding data and reportedly matched Anthropic Claude Opus performance in internal tests at SpaceX and Tesla.
• First SpaceXAI-Cursor joint model, this release tests whether a coding-tool company with new compute access can reduce dependence on third-party frontier models while retaining developer trust.

Microsoft is already routing tens of thousands of weekly AI prompts in Excel, Outlook, and GitHub Copilot to its in-house MAI models, with CEO Mustafa Suleiman stating the explicit goal is to eliminate payments to OpenAI and Anthropic.

• MAI-Transcribe-1.5 is live in Teams, MAI-Image-2 is in testing, and the migration is confirmed by Bloomberg, TechCrunch, and The Decoder as of July 7.
• Strategic implication for enterprise buyers: Microsoft's $69B Activision-scale dependency on OpenAI is unwinding in production, not on a roadmap, which reshapes how enterprises should think about Copilot's model provenance.

Alstom, a rail manufacturer with 88,000 employees and €19.2B in 2025 revenue, expanded its decade-old Zscaler Zero Trust Exchange deployment to inspect and govern LLM data flows across its enterprise.

• Existing ZTNA infrastructure was adapted rather than replaced, with Alstom extending the same identity-verified access controls it uses for applications to cover data moving into and out of large language models.
• The core risk addressed is that GenAI creates new egress paths for sensitive data that bypass traditional DLP policies, requiring runtime visibility into what employees send to models.

Chainguard's Athena coalition processed more than 40,000 open source vulnerability findings in three weeks as AI-powered exploit discovery compresses the window between vulnerability discovery and active exploitation.

• Eight new members joined including Akamai, JFrog, Morgan Stanley, and Qualys, reflecting both security vendors and a major financial institution treating coordinated OSS defense as a shared infrastructure problem.
• The threat model has changed: frontier AI models can now identify exploitable weaknesses at scale, which is the same capability that triggered the US government's hold on GPT-5.6, creating a race between AI-assisted attack and AI-assisted defense.