What happened
A default-setting change Anthropic announced the prior week took effect on August 14: auto mode is now the default for Claude Code, the company's AI coding agent, across Pro, Max, and Team accounts. In auto mode, the agent no longer asks approval for each individual action; it proceeds on its own unless an action is judged irreversible, destructive, or aimed outside the working environment. Anthropic's stated rationale is that click-through approvals were not actually protecting anyone: in a company study of 1,053 paid users, human review caught 13.6 percent of harmful actions, while auto mode's automated safeguards caught 89 percent. The company also notes users approve 97 percent of permission prompts, classic habituation. The new default ships alongside prompt-injection screening and configurable hard deny rules that block specific actions outright, and Anthropic continued shipping updates through the week, including GitLab merge request support on August 15.
Why it matters for your business
AI coding agents are no longer a big-company tool; small firms use them to maintain websites, write scripts, and glue systems together, sometimes through a contractor you have never asked about this. A default change means behavior shifted for existing users who did nothing, which is exactly when settings deserve a review. Anthropic's habituation data is worth absorbing on its own: if your safety process is a human clicking approve dozens of times a day, it likely approves everything. That applies to expense approvals and access requests as much as to code.
What to do about it
- If anyone on your team uses Claude Code or a similar agent, have them confirm which mode it now runs in and what it can reach.
- Set explicit deny rules for the things that must never happen automatically: touching production systems, secrets files, or customer data.
- Give agents their own sandboxed accounts and repositories, so autonomy is bounded by access, not by trust in prompts.