Auto mode is now the default in Claude Code for Pro, Max, and Team plans

Claude Code now defaults to its "Auto mode" starting August 14th for Pro, Max, and Team plans, according to Anthropic. This decision stems from Anthropic's confidence in the safety of Auto mode, as discussed in their recent Fireside Chat. Anthropic's CEO, Cat Wu, stated that almost everyone within the company uses Auto mode and asserted that they have mitigated all major risks, including prompt injection and data exfiltration, better than human reviewers.

In a test involving 1,053 paid testers, only 13.6% of humans would approve a harmful action, whereas Auto mode would have blocked 89% of such requests. However, 11% of cases still slip through the cracks. Anthropic commissioned an evaluation from third-party firm Trajectory Labs, which tested various models, including Claude Code, against 72 indirect prompt injection scenarios.

The evaluation found no successful attacks against Auto mode versions of Claude Fable 5, Opus 5, and Sonnet 5. Despite these assurances, concerns remain about accidental damaging actions, such as deleting crucial files or clearing databases, and more specifically, prompt injection scenarios where malicious instructions are hidden within seemingly harmless content.

One potential threat highlighted is a malicious third-party package that could exfiltrate data if unknowingly run. While Anthropic claims to have effectively addressed these issues, it is recommended to seek more independent confirmation of their findings.