Anthropic to Watermark All Claude-Generated Text Under EU AI Act Rules
Anthropic confirmed it will embed imperceptible watermarks into text produced by Claude models and attach digitally signed provenance metadata to generated files. The move complies with the EU AI Act’s transparency obligations, which became enforceable on August 2, and with the EU’s Code of Practice on Transparency of AI-Generated Content that Anthropic signed alongside roughly 190 other organizations.
The key detail: these markings apply worldwide, not only to EU users, and the watermark is designed to survive copy-and-paste. Text gets an invisible embedded signal readers never see; files use the open C2PA standard for signed provenance. Coverage starts with every Claude model launched in the EU on or after August 2, and Anthropic says it’s working to extend marking to older models. Skeptics note watermarks can be stripped or paraphrased away — but as a default, it makes AI text far easier to trace.
Claude Code Auto Mode Becomes the Default on August 14
Anthropic is flipping Claude Code’s default permission mode to auto for Pro, Max, and Team plans starting August 14. Instead of stopping to ask for manual approval on each action, a classifier screens every tool call and lets Claude proceed — pausing only when an action is judged irreversible, destructive, or aimed outside your environment.
It’s a meaningful shift in defaults: less friction for the common case, but it puts more weight on the classifier’s judgment about what counts as risky. Teams that prefer the old behavior can still switch back to manual approvals, and the change follows auto mode already becoming the default on Bedrock, Vertex AI, and Foundry.
MCP 2026-07-28 Spec Arrives in Claude; Directory Passes 950 Servers
Claude now supports the MCP 2026-07-28 spec, one of the protocol’s most significant releases — it moves from a bidirectional stateful model to a stateless request/response core, letting servers deploy on serverless and edge infrastructure. The update brings stronger OAuth and OIDC authorization plus versioned extensions for Apps and Tasks, so developers can add interactive UIs and long-running work without touching the core protocol.
New connector capabilities ride along: embedded UI, enterprise-managed auth, observability, and private network tunnels. Claude now lists over 950 MCP servers in its connectors directory, used by millions of people daily, and MCP recently crossed 400M monthly SDK downloads — a 4x jump this year that cements it as the default way to wire agents to apps.
Workbench and Prompt Tools APIs Retire August 17; Code Execution Gains REPL State
Two housekeeping notes for builders. The experimental prompt tools APIs and the Workbench are being retired on August 17, so anyone relying on those flows should migrate now. On the upside, the Python, TypeScript, Go, Java, Ruby, PHP, and C# SDKs now support code_execution_20260120 — the code execution tool version that adds REPL state persistence.
Persistent REPL state is the quietly useful part: variables, imports, and intermediate results survive across calls within a session, so multi-step data work and iterative debugging no longer restart from scratch each turn. It’s the kind of change that makes agentic, tool-heavy workflows feel far less brittle.
Cowork Reaches Web and Mobile — and the Data Shows Most Users Aren’t Coding
As Anthropic rolls Claude Cowork out to the web at claude.ai and to iOS and Android, its own usage data tells a notable story: most Cowork users aren’t writing code. The surfaces run on remote sessions hosted on Anthropic’s servers, saved to your account, able to keep working and run scheduled tasks with no device online.
That mix — document work, research, spreadsheets, file wrangling — is the tell. Cowork is aimed less at engineers and more at the broader knowledge-work crowd, and unifying chat and Cowork onto one home screen (with projects and artifacts persisting across both) is Anthropic’s bet that “agent that does the work” belongs everywhere, not just in a terminal.
Claude for Government Enters Beta
Anthropic opened Claude for Government in beta, a packaging aimed at public-sector teams that need Claude’s capabilities inside the compliance and controls their agencies require. It slots alongside the enterprise push — richer admin analytics, model-level entitlements, and spend alerts for Claude Enterprise — that gives administrators deeper visibility into usage, cost, and productivity.
The through-line with today’s watermarking news is hard to miss: Anthropic is leaning into the parts of the market where provenance, auditability, and governance are non-negotiable. Government and regulated enterprise are exactly the buyers who care most about where AI output came from and who approved each action.
Compliance Is Becoming a Feature, Not a Cost
Watermarking every Claude output worldwide is the headline, but look at what surrounds it. Auto mode ships with a classifier that knows what it’s not allowed to do. Claude for Government arrives with governance baked in. Enterprise admins get spend alerts and entitlement controls. Anthropic could have treated the EU AI Act as a regional checkbox — instead it’s applying the transparency machinery globally and turning it into a selling point.
That’s the strategic read. As agents take more autonomous action, the buyers who matter most — governments, banks, defense, healthcare — don’t just ask “is it capable?” They ask “can I prove what it did, and where the output came from?” By making provenance and controls the default rather than an enterprise upsell, Anthropic is betting that trust infrastructure is what wins the accounts capability alone can’t. Regulation, handled this way, stops being a tax and starts being a moat.