Back to rankings

Claude Code Asks Before Sending Your Account Email to the Network and Warns About Unrequested Cloud Task Connectors

Claude CodeView original changelog

Claude Code 2.1.317 adds a check that asks for approval before a web or shell tool call sends the user's account email address to the network. It also warns when a new cloud scheduled task is given connectors the user did not request. Agent SDK hook callbacks can now fail closed with onFailure: "block", and auto mode now says so when it has no verdict for a subagent's tool call.

Key Takeaways

  • Account email is now gated: a web or shell tool call that would send the signed-in email address to the network triggers an approval prompt first.
  • Cloud scheduled tasks get a connector warning, so connectors the user did not ask for are flagged at /schedule time with instructions to remove them.
  • Agent SDK hooks can fail closed with onFailure: "block", so a broken callback stops the action instead of allowing it.
  • Auto mode now explains silence: when its check cannot rule on a subagent's tool call, a notice appears in the main view.
  • Unattended jobs benefit most, since scheduled tasks and subagents run without a human watching each step.
  • All four changes target the same gap, the distance between what the agent can do and what the user authorized.

Guarding what leaves the machine

Claude Code 2.1.317 adds several controls that keep the agent within what the user actually authorized. The headline change is a new check that pauses and asks before a web or shell tool call sends the user's account email address to the network. Previously a command or request that embedded the signed-in email could go out without any specific prompt. The new check puts that decision in front of the user, which matters for anyone running Claude Code on shared machines, in regulated environments, or against untrusted prompts and repositories.

A warning for cloud scheduled tasks

When a user creates a new cloud scheduled task with /schedule, the task can end up with connectors the user never asked for. Claude Code 2.1.317 now shows a warning in that case and explains how to remove the extra connectors. Scheduled tasks run unattended, so an unexpected connector means an unattended job that can reach a service the user did not intend to grant it. Surfacing it at creation time lets users trim the task before it runs.

Hook callbacks that fail closed

For developers building on the Agent SDK, hook callbacks gain an onFailure: "block" option. With it, a callback that errors or fails blocks the action it was guarding instead of letting the action proceed. This turns a policy hook from fail-open into fail-closed, which is the safer default for approval and compliance checks.

Clearer auto mode feedback for subagents

Auto mode now shows a notice when its safety check has no answer for a subagent's tool call. In earlier versions that situation left nothing visible in the main view, so a user could not tell why a subagent stalled or whether a call had been vetted. The notice makes the gap visible.

Why it matters

Each of these changes narrows the distance between what the agent can do and what the user explicitly approved: an outbound identifier, a background task's connectors, a failing policy hook, and an unreviewed subagent call.