Claude Apps: Chrome GA, Memory Everywhere, Watermarking, Teachers

Analysis

Claude in Chrome is generally available

Claude in Chrome moved out of limited testing and is now generally available on every paid Claude plan. It performs autonomous browser actions, gated by a safety classifier, and Enterprise administrators can restrict it with domain allowlists.

The governance framing is straightforward: an agent operating inside a logged-in browser session inherits whatever that session can do. If your organization has a paid plan, the capability is present unless an admin scopes it. Deciding after users have adopted it is harder than deciding now.

• Availability follows the paid plan, so treat it as opt-out rather than opt-in.
• Configure Enterprise domain allowlists before broad rollout, not after.
• Write user guidance for what should never be done in an agent-controlled tab.
• Pair browser automation with existing session and SSO risk reviews.

Analysis

Memory is now one system with user controls

Chat memory and Cowork memory are the same memory. Users get topic-level edit and delete, sensitive topics are off by default, and the plan-level defaults differ in a way that matters: on by default for Free, Pro, and Max; admin-gated and off for Team and Enterprise until the user enables it.

That split is the single most common misreading of this change. On managed plans, memory does not switch itself on. Admins asking "is Claude retaining our conversations?" need the plan-specific answer, and users on Team plans asking "why doesn’t Claude remember me?" usually need an admin to enable the feature rather than a bug report.

• Confirm the memory setting for your own workspace rather than assuming the consumer default.
• Tell users that topic-level edit and delete exist; it is the answer to most memory complaints.
• Note that sensitive topics are excluded by default.
• Fold memory into the same data-handling review you use for connectors.

Analysis

Generated text now carries a watermark

Anthropic announced text watermarking on August 14, 2026, framed against the EU AI Act and its Code of Practice. Future Claude models watermark generated text. The announcement is precise about what the watermark is not: no practical quality impact, no hidden characters, no extra tokens, and not traceable to a person, organization, or conversation. It was global at launch, a detection API is described as coming, and models published before August 2, 2026 have a transition period.

For content and legal teams, the useful reading is that this is a provenance signal about model output, not an identity signal about an author. Policies that treat a watermark as attribution to a specific user are misreading the feature.

Analysis

Claude for Teachers and the education surface

Claude for Teachers launched on July 14, 2026 for verified US K-12 educators: free premium access, teaching skills, and curricula mapped to all 50 states, with a one-year sign-up window. A school and district offering followed on August 28, 2026.

The privacy detail is the one to carry into procurement conversations: the offering is FERPA-configured, and identifiable student records require school or district authorization. An individual teacher signing up is not the same authorization as a district agreement.

Analysis

Files, connectors, and Slack still need owners

The workflow features that predate this issue have not changed shape. Claude support documents file creation and editing across spreadsheets, presentations, documents, PDFs, analysis outputs, scripts, and visualizations, which makes Claude a work-product generator rather than only an answer generator. Connectors remain MCP-based and inherit the permission model of the connected service.

Claude Tag remains the Slack path for Team and Enterprise, announced as a beta on June 23, 2026. Verify workspace access rather than assuming it, and pick pilot channels before opening it everywhere.

• Review which connector sources are enabled per team and whether permissions inherit or need policy.
• Decide what "final" means for a file that came out of a chat: reviewed, editable, attributable, stored where.
• Keep sensitive-project channels and connector policy aligned.
• Route each app release note to an owner: user education, admin policy, engineering, security, or support.