Claude Code 2.1.325 and Codex CLI 0.169.0 were released on the same day, roughly 12 hours apart, and both tightened default permissions for unattended use. This was not an ordinary update: automated jobs may suddenly wait for confirmation, while configurations in untrusted projects may be skipped entirely.
What This Is
Claude Code has two major changes: once sandboxing is enabled, Bash commands are no longer automatically allowed by default; and scheduled tasks and /loop runs involving skills with allowed-tools or hooks now require confirmation before proceeding. To restore the previous behavior, explicitly set sandbox.autoAllowBashIfSandboxed: true and configure permissions.allow according to the documentation.
Codex, meanwhile, restricts project-level configurations in untrusted directories, including hooks and rules under .codex/ and project-defined MCP servers—a standard interface that lets external tools and services connect to AI workflows. To restore the relevant capabilities, add --trust to codex exec or codex review, or mark the project as trusted in the user-level configuration.
Notably, neither company explained in its release notes why it adjusted the defaults. The official documentation also warns that sandboxing only reduces risk and does not provide complete isolation; Claude Code's default sandbox may still be able to read some credential files. Treating “sandboxed” as “completely safe” therefore remains too crude an assumption.
Industry View
We believe this adjustment sends a clear signal: AI programming tools are shifting from “execute as much as possible on your behalf” toward “first prove that this code and these permissions are trustworthy.” For enterprises that frequently handle code, credentials, and production environments, one fewer automatic approval brings clearer operational boundaries and accountability. This is especially true for scheduled tasks, recurring execution, and project configurations, where a mistake often has consequences beyond a single conversation.
But the opposing argument is equally valid. Tighter defaults directly disrupt existing pipelines, forcing teams to re-audit every permission, add trust markers, and maintain configurations across versions. For individual developers or small teams, this may amount to little more than post-upgrade friction; for enterprises that require stable unattended operation, it may become a deployment cost.
More concerning is that if enterprises simply restore automatic approvals to regain efficiency, the security benefits may disappear with them. The sensible approach is not to grant blanket access, but to apply tiered authorization by directory, command, and task type, while retaining logging, alerts, and revocation mechanisms.
Impact on Regular People
For enterprise IT: Permission inventories, trusted directories, and sandbox policies need to be brought under version control. Before upgrading, teams should also verify that scheduled tasks, code review, and deployment workflows can still run continuously.
For individual professionals: Employees using Claude Code or Codex will find that some automations have suddenly stopped. This does not mean the tasks have failed; the tools are requiring risks to be reconfirmed. Understanding the differences between configurations will directly affect productivity.
For the consumer market: Consumers may not see any obvious changes in the short term, but enterprise AI services will place greater emphasis on controllability and auditability. Products built around “fully automated delegation” may add confirmation steps or raise the barrier to use.