Claude in Slack can now use your own connectors, not just admin ones

Scoped to requests you make, with review or auto posting — plus GitHub fail-closed local sandboxes for Copilot agents and new code-review controls.

Nowline SEP 25 12:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Bring your own connectors into a Slack channel

    Claude Tag can now use your personal connectors — calendar, cloud storage, CRM, deploy tools — on requests you make in a channel, not just the tools an admin wired up. Pick Review mode (you approve before it posts) or Auto mode (it posts unless it flags something sensitive); it's rolling out on Team plans now, Enterprise next. Catch: it only applies to requests you initiate — scheduled and unattended tasks still run on the admin's channel-level connectors.

  • Copilot can now sandbox local agent sessions

    The GitHub Copilot app got fail-closed local sandboxing in public preview: per-project policies over the filesystem (allow/deny folder lists), network (outbound and local), and your git/gh credentials. Fail-closed means if your OS can’t enforce the policy the shell errors out instead of running unsandboxed. Turn on "Sandbox new sessions" in project settings or run /sandbox on; enterprise settings can only tighten, never loosen, the project defaults. Applies to local repo sessions, not cloud or remote hosts.

  • Copilot code review gets personal + org controls

    Personal review controls are now generally available, and enterprise default effort settings rolled out across the expanded Copilot plans — so you can set how you request a review and how deep Copilot goes on each PR, and orgs can standardize it. Small on paper, but it changes the default behavior of the reviewer bot that comments on your pull requests.

  • Mistral Common v1.12.0 tightens validation

    If you build on Mistral, the tokenizer/validation library added an agnostic validation mode, safer multimodal (image and audio) handling, and now allows null assistant content in OpenAI-format requests. It also deprecates Tokenized.text in favor of tokenizer.decode and fixes an image-URL loading crash. Maintenance-grade, but the deprecation is worth a look before you bump.