Digital Colliers Daily Briefing — July 8, 2026
The agentic-coworker race broadened well past the terminal on Tuesday, as Anthropic pushed Claude Cowork onto phones and browsers with cloud execution — a direct challenge to Meta, Google, and OpenAI for control of everyday knowledge work. Meta answered on a different front, unveiling its first Superintelligence Labs image model, Muse Image, with an opt-out policy that lets any Instagram user pull public accounts into generated pictures. And the geopolitical layer got louder: China's CNVD accused Claude Code of shipping "backdoor vulnerabilities" as US lawmakers weighed limits on domestic use of Chinese models and DeepSeek moved toward designing its own silicon.
1. Claude Cowork Goes Multi-Device, With Tasks That Run While the Laptop Is Closed
What happened. Anthropic on Tuesday extended Claude Cowork — the general-purpose sibling of Claude Code launched as a desktop app in January — to iOS, Android, and the web, starting with Max subscribers ($100/month) and rolling out to Pro ($20/month) "in the coming weeks," according to The Verge and Wired. The important architectural change is that Cowork sessions now run in the cloud by default. Previously, users had to leave their laptops open for scheduled tasks to execute via the Dispatch feature; now, per Anthropic's launch video, a task like "prep Monday's client meeting" can pull email threads, Slack channels, transcripts, and news overnight and surface a draft brief and unsent follow-up email the next morning. Anthropic also unified Chat and Cowork on web and desktop, with projects and artifacts shared across surfaces. The desktop app remains the venue for deep work with local-file and browser access.
Why it matters. Anthropic paired the launch with usage data from 1.2 million anonymized Cowork sessions across 600,000+ organizations in late May. Business process operations — reports, onboarding checklists, spreadsheet reconciliation — accounted for 33.4% of use; content and copywriting, 16.4%; software development, only 8.7%. As TechCrunch framed it, the coding-agent wars are spilling into the rest of the office. Success now hinges less on chatbot benchmark scores than on owning the surfaces where finance, HR, marketing, and admin work actually happens. Latent Space's roundup noted the same convergence at OpenAI (Codex Mobile, Codex Remote) and Google (Gemini API Managed Agents added background execution and remote MCP servers this week).
Who is affected. Max subscribers get the beta first; Pro users next; free-tier access remains unclear. Enterprises evaluating agent platforms now have a concrete comparison point against OpenAI Codex and Google Spark. IT and security teams inherit new questions about cloud-executed agents acting on behalf of users across devices — including the prompt-injection exposure Wired's reviewer flagged from her January testing.
What to watch next. Pricing and rate-limit tiers as Cowork descends to Pro; whether Claude Tag (the always-on Slack teammate) and Cowork converge; and how Anthropic handles the audit-trail and consent flows enterprises will demand once background agents are drafting outbound email unattended.
Sources:
- Claude Cowork expands to mobile and web — TechCrunch AI
- Shut Those Laptops! Anthropic Puts Its Claude Cowork Agent on Your Phone — Wired
- Claude Cowork: coming to mobile and web — YouTube · Claude
- Anthropic is launching Claude Cowork on mobile and web — The Verge AI
- [AINews] Lilian Weng summarizes 35 papers on Harness Engineering for RSI — Latent Space
2. Muse Image Ships Across Meta's Apps — With Public Instagram Photos Opted In by Default
What happened. Meta Superintelligence Labs released Muse Image, its first in-house image generator, replacing Llama-based image tooling and slotting into the Meta AI app, Instagram, and WhatsApp, with Facebook and Messenger to follow. Alexandr Wang, who runs the lab, described Muse Image on Threads as "agentic" — it pairs with the Muse Spark LLM to reason, search the web, and plan before rendering, per The Verge. On the Image Arena leaderboard it debuted at #2, behind GPT Image 2. A Muse Video model is in development.
The controversy is the tagging behavior. Any user can tag a public Instagram account in a prompt and generate AI images using that person's likeness. Meta frames it as a way to build custom invitations or collaborative concepts. Public accounts are opted in by default; to opt out, users must navigate to Instagram's Sharing and reuse settings and toggle off "Allow people to use your content on Instagram and with AI features on Meta" for both Posts and Reels. Wired reported that as of Tuesday afternoon the updated language had not yet appeared in every account's settings. Meta's help page confirms users will not be notified when their content is used, and pre-existing generations will not be deleted if a user later opts out or switches to private.
Why it matters. This is the second major default-opt-in AI-training pattern of the past year — Google Search now retains reverse-image-search uploads for training — and it lands on a platform with billions of accounts and a documented history of consent failures. TechCrunch noted the throughline: the $5 billion FTC fine in 2019 over Cambridge Analytica, and the 2021 shutdown of Facebook's facial-recognition system amid biometric-data litigation. Regulators in the EU and UK, which have been aggressive on generative-AI training consent, are likely to treat "tag any public profile to synthesize their likeness" as a distinct harm from bulk training-data scraping.
Who is affected. Every public Instagram user, immediately. Creators and public figures face the sharpest exposure — impersonation, non-consensual imagery, and reputational risk with no notification. Brands and marketers get a new low-friction ad-mockup pipeline and a Facebook Marketplace integration (visualizing secondhand furniture in one's own space was Meta's showcase use).
What to watch next. Regulatory response from Ireland's DPC and the UK ICO; whether Meta adds opt-in defaults or notification for tagged users under pressure; how quickly abuse cases — deepfake harassment, non-consensual imagery — surface; and Muse Video's release timeline, which will multiply every issue above.
Sources:
- Meta Now Lets Anyone Use Your Instagram Photos in AI Images—Unless You Opt Out — Wired
- Meta's new Muse Image model can pull other Instagram users into AI photos — The Verge AI
- Meta just launched a new AI generator, Muse Image, and users are already pushing back over use of their photos — TechCrunch AI
3. CNVD Flags Claude Code as a "Backdoor" as Washington Weighs Curbs on Chinese Models
What happened. China's National Vulnerability Database (CNVD) issued an advisory alleging "security backdoor vulnerabilities" in Anthropic's Claude Code that transmit "sensitive information" including "location and identity to remote servers," according to Raffaele Huang's reporting for the Wall Street Journal. CNVD advised users to uninstall the software or update to the latest version. The advisory arrives the same week CNBC reports a US House committee investigation is examining domestic corporate adoption of Chinese AI models, with lawmakers weighing curbs. Separately, Reuters reported (via Ars Technica) that DeepSeek — facing US export controls on advanced accelerators — has spent about a year quietly working toward its own silicon, meeting hardware partners and hiring engineers.
Why it matters. The "backdoor" framing mirrors the language US authorities have used against Chinese vendors from Huawei to TikTok, and its deployment against a flagship American AI coding tool signals that the vulnerability-disclosure apparatus is now a lever in the AI decoupling. Anthropic already restricts sales in China; the CNVD notice functionally hardens that separation from the other side. Combined with the House investigation and DeepSeek's silicon push, the market is bifurcating along the full stack — models, developer tooling, and chips.
Who is affected. Multinational engineering teams using Claude Code inside China now face compliance ambiguity and likely internal bans. US enterprises quietly running DeepSeek, Qwen, or other Chinese open-weight models — a practice that has grown as those models close performance gaps — face potential regulatory constraints. Nvidia and other US chip vendors watch DeepSeek's silicon plans as a demand-side signal, however distant commercial viability may be.
What to watch next. Anthropic's technical response to CNVD's specific claims; the scope of any House committee recommendations, particularly whether they target federal contractors or reach further; and whether DeepSeek's chip effort attracts SMIC or other domestic foundry partners under continued export-control tightening.
Sources:
- China's CNVD says it found "security backdoor vulnerabilities" in Claude Code that "send sensitive information" like "location and identity to remote servers" (Raffaele Huang/Wall Street Journal) — Techmeme
- US lawmakers are considering strategies to curb US companies' growing adoption of Chinese AI models, as a House committee investigation investigates the risks (Kai Nicol-Schwarz/CNBC) — Techmeme
- Facing US export controls, China's DeepSeek plans to make its own chips — Ars Technica
Tuesday's three stories describe the same underlying shift from different angles: AI is moving off the chatbot surface and into the ambient plumbing of work, social media, and geopolitics. Anthropic wants agents that run without the user present; Meta wants an image model that can conscript any public profile into a prompt; Beijing and Washington want the plumbing itself split down national lines. The next round of fights — over consent defaults, background-agent auditability, and cross-border tool bans — will be decided less by model quality than by who sets the rules of the surfaces these systems now inhabit.