CONFIRMED. Anthropic released Claude Code 2.1.325 with security fixes and an important change for automation: commands running inside its sandbox now follow the normal permission flow by default. The release also fixes background sessions that used and billed a programming key even though they had started with a claude.ai sign-in. For businesses, updating can bring stronger control, but also jobs that begin waiting for approval.
An update that changes operations
The official release notes were published on November 13. Claude Code is the artificial intelligence assistant that works in a terminal and a project’s files. This version is not a smarter underlying model. It revises how the assistant performs work, respects boundaries and preserves its sign-in identity.
Independent tracker Claude Code Changelog had flagged the arrival of 2.1.325 before the official release appeared. That provides distribution context, not independent validation of the fixes. Anthropic is the source for the description of the corrected behavior.
Isolation no longer means automatic approval
The most consequential change for automated workflows concerns Bash commands, which operate the system. Even within the sandbox, an isolated environment restricting access, they now follow the normal permission process. The notes identify sandbox.autoAllowBashIfSandboxed as the option for organizations that want to retain execution without prompts. That is not a reason to enable it indiscriminately.
Scheduled tasks and the /loop command also ask before running skills that declare allowed tools or supporting hooks. For a small Brazilian agency leaving an agent to prepare changes overnight, the practical consequence could be a job waiting for a person instead of a finished deliverable in the morning. Specific permissions deserve review before the workflow is relied upon again.
The wrong credential can mean the wrong bill
Anthropic says it corrected background sessions started with a claude.ai sign-in that instead used an application programming interface, or API, key or token inherited from the shell that started the service. The notes explicitly say that path generated billing. They do not specify how many users were affected, amounts, refunds or any change in prices.
Another correction prevents token counting from sending a conversation to Amazon Bedrock when the provider policy does not allow it. Tokens are units of content processed by models. For businesses handling client projects, the issue concerns both expenditure and the destination of data.
Update without mistaking fixes for guarantees
The release includes a security issue in decoding zstd-compressed responses, alongside problems involving deny rules, compound commands and configuration protection. The announcement does not detail exploitation or establish that a breach occurred. It also fixes resumed sessions losing recent messages and context errors on third-party gateways.
This page follows Claude Code’s evolution and retains the context of version 2.1.323, which introduced experimental Code Mode, where Claude writes JavaScript to invoke tools. Version 2.1.325 corrects behavior in that mode, including work in the wrong folder and resumed sessions starting execution without a new request. Our assessment is to prioritize the update, followed by a short check of existing workflows. Stronger security and operational continuity must be evaluated together. Release notes alone do not demonstrate savings or justify promising unsupervised autonomy.
Sources
Anthropic: Claude Code 2.1.325; Claude Code Changelog: 2.1.325; Anthropic: Code Mode, 2.1.323.
