Claude gets invisible watermarks, self-hosted runners go beta, Compliance API covers Cowork and Code

Anthropic announced on August 11 that all Claude models released after August 2 automatically embed an imperceptible watermark in generated text to comply with the EU AI Act's Transparency Code, which took effect August 2. The watermark does not change text meaning or readability, but travels with text when copied and pasted and may persist through light editing — though heavy editing can erase it. For files, Anthropic is using the C2PA open standard. Older models are being updated to add watermarking support on a rolling basis. The policy applies globally, not just in the EU. Anthropic notes that the presence of a watermark does not guarantee AI use and its absence does not confirm human authorship. Google already watermarks Gemini text; OpenAI has not yet announced its approach.

The Claude in Chrome browser extension's side panel has been renamed Claude Cowork, unifying the branding across desktop app, web, mobile, and browser contexts. The extension still requires the Claude desktop app for file access and computer use. Claude in Chrome does not yet run on other Chromium-based browsers or on mobile. Separately, Claude's connectors directory now lists over 950 MCP servers, used by millions of people every day — up from roughly 800 at the MCP 2026-07-28 launch.

Anthropic opened the public beta of self-hosted environments for Claude Code on August 6 with Claude Code v2.1.224. A single command — claude self-hosted-runner — turns your own machines or containers into compute for Claude Code sessions started from web, mobile, desktop, or scheduled routines. Sessions run inside your network next to your internal services, toolchains, and security controls, rather than on Anthropic infrastructure. Two runner modes: Fixed (a set number always running, sessions distributed across them) and On-Demand (an orchestrator starts runners as sessions arrive and stops them when done). Available on Team and Enterprise plans, disabled by default. Anthropic recommends hosted environments for most teams and says self-hosted is for organizations with strict data-residency, network-isolation, or compliance requirements — it requires engineering ownership of runner images, updates, and orchestrator operation.

The Claude Compliance API now covers Cowork sessions across desktop, web, and mobile, and Claude Code sessions in the CLI and desktop app. Security and compliance teams can pull unified session content and metadata for audits and eDiscovery through the same API and existing Compliance Access Key — no new integration required. Coverage is in beta for Claude Enterprise customers. Previously, Compliance API coverage was limited to Claude chat sessions.

Enterprise plans can now enable skill and plugin security scanning to automatically check third-party skills and plugins for malicious content when someone uploads or edits them. Enable in Enterprise admin settings under Security. Alongside this, inference hooks entered beta for Enterprise — giving compliance teams real-time DLP enforcement by inspecting and optionally blocking requests and responses before they reach users. Both features address the growing enterprise concern around supply-chain attacks on Claude Code skills and plugins.

Anthropic confirmed it is building an in-house silicon team to design custom AI chips — joining OpenAI, Google, Meta, and Amazon in vertically integrating compute. The team is in early stages and no chip timeline or architecture details were disclosed. The move comes as Anthropic negotiates a second $36B chip financing deal for Google TPUs, suggesting the in-house silicon effort is a long-term hedge rather than a near-term compute replacement.

A dense Claude Code release week: subagent forking is now on by default — a subagent_type: 'fork' subagent inherits the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions now run in the background by default. Type @ in the prompt to mention another live Claude session by name; Claude uses SendMessage to reach that session directly, and SendMessage now delivers to a bare name matching one live session without asking to confirm. GitLab merge request support landed. VS Code gained sidebar session groups for organizing multiple active sessions. Remote Control gained --continue for resuming the most recent session. MCP OAuth sign-in failures caused by redirect URI mismatches for pre-registered clients like Slack were fixed.

Claude Code v2.1.224 added archive plugin source — install plugins from a ZIP over HTTPS without git or npm, with optional SHA-256 pinning for integrity verification. New crossSessionInbound setting holds cross-session messages sent to sessions running with bypassed permissions for your manual approval. New dialogExpiry setting auto-delivers messages to other sessions. ANTHROPIC_BEDROCK_REGION_PREFIX env var lets Bedrock users prefer a specific cross-region inference profile over the AWS_REGION-derived default — useful for teams pinning to specific regional endpoints.

The temporary 50% Claude Code weekly usage limit boost — in place since May 13, 2026 and extended multiple times — ends on August 19, 2026. From August 20, weekly limits return to their standard levels for Pro, Max, Team, and Enterprise plans. If your team has built workflows that depend on the elevated limits, audit your typical weekly usage now and identify which tasks to prioritize within tighter constraints or route to the API.