Anthropic released Claude Code 2.1.317 on 3 November 2026. Most of the notes are fixes, but two additions matter to anyone commissioning AI-assisted development, because both are about what an AI agent can send and reach without a person noticing.

What the changelog says

• Email address check. Claude Code now asks before a web or shell tool call sends your account email address to the network.
• Scheduled task warning. When a new cloud scheduled task is created with connectors you did not ask for, /schedule now warns you and explains how to remove them.
• Failed hooks can block. A new onFailure: "block" option for Agent SDK hook callbacks means that if a safety check fails, the action it was guarding is stopped rather than allowed through.
• Cost and identity reporting. Gateway administrators get a cost report endpoint, plan and seat details in telemetry, and daily records of each developer’s identity-provider groups.
• Reliability. The release lists 62 bug fixes, including claude -p now exiting with an error when a full disk loses its transcript, instead of reporting success.

What it means

Each item is a small guardrail on automation that runs unattended. Scheduled agents are the sensitive case: they act while nobody is watching, so the default list of services they can touch matters. A warning at creation time is a sign that vendors expect buyers to ask what their agents can reach.

So what

If your supplier uses AI agents or scheduled automation on your code, ask which connectors and accounts each one can use, who approved them, and what stops a failed safety check from letting work through. Our AI-assisted development service documents those permissions for every client. To discuss your setup, start a project.

← All research