Product news: ChatGPT adds reservations while Claude scans plugins, making purpose and permission part of AI literacy

This product-news report follows a practical shift from AI that recommends an answer to AI that can initiate an external action. OpenAI's August 10 ChatGPT update can surface restaurant availability and open a reservation flow. Anthropic's August enterprise beta scans newly uploaded or edited third-party skills and plugins. Google frames Gemini for Education around three purposes—teach, learn and work—inside an institutionally managed product. Together, the releases make the line between a helpful suggestion and an authorized action visible.

OpenAI says users can specify a location, time, party size and preferences, then refine results before selecting a reservation. OpenTable availability is offered globally, Resy in the United States, and Yelp in the United States and Canada, subject to provider coverage. If a booking needs a name or email address, ChatGPT asks permission before filling those fields. Availability can change, so users must confirm the details. ChatGPT does not retain the completed reservation; later changes and cancellations go through the booking provider.

The account boundary matters for education. OpenAI lists the feature across consumer ChatGPT plans using GPT-5.6 Sol and GPT-5.6 Luna, but not in ChatGPT Work or in Enterprise, Edu and Business workspaces. A personal-account capability should not be assumed to exist in an institutionally governed account. Although booking is not an education product, it gives students a concrete agentic-action case: a model proposes options, a third party supplies changing availability, personal information crosses a boundary with consent, and the provider holds the final record.

Anthropic's beta addresses another link in that chain. Enterprise owners can enable scanning for new or edited third-party skills and plugins in Claude, Claude Cowork and enterprise marketplaces. Most scans take about one to two minutes and return pass, warn or fail. A warning requires acknowledgement; a failure blocks installation. The feature is off by default, and a failed item must be changed and resubmitted rather than overridden.

The scan is a risk signal, not a quality certificate. It does not retroactively cover existing items, Claude-created skills, MCP-delivered skills, MCP servers or hooks, and it is unavailable in some customer-managed encryption, zero-data-retention and HIPAA configurations. A pass does not establish that a plugin is pedagogically accurate, accessible, privacy-minimizing or suitable for a learner. Schools still need a trusted source, permissions review and independent evaluation.

Gemini for Education offers a purpose test before an action-capable tool is adopted. Google groups examples under teaching, learning and work: lesson and assessment design; explanations, practice and writing feedback; and communications, administration and research. It also describes administrator controls, a distinct under-18 experience, and institutional data that is not human-reviewed or used to train models. These are vendor-stated conditions, not evidence of learning gains, but they connect capability to purpose and account context.

Hong Kong schools and universities can keep an action ledger for pilots: the teaching, learning or work purpose; account and plan; external service and data fields; user permission; human confirmation; authoritative receipt; and correction or cancellation route. Third-party skills should be tested with synthetic data even after a scan passes. Students can compare what an assistant proposed with what the external system recorded and identify where responsibility changed hands. AI literacy now includes understanding what was authorized, what crossed a boundary and who can undo the result.