Anthropic shifts Claude Code to auto mode by default, citing superior safety metrics
The company argues that automated oversight outperforms manual review, which suffered from high user habituation rates in internal testing.
Anthropic has announced that Claude Code’s auto mode will become the default setting for Pro, Max, and Team accounts commencing 14 August 2026. This update marks a significant shift in the software’s operational model, allowing the artificial intelligence agent to proceed with programming tasks without seeking human approval for every individual step. The system will only intervene if an action is classified as irreversible, destructive, or directed outside the user’s immediate environment.
The decision to automate the workflow is underpinned by internal testing data provided by the company. In a study involving 1,053 paid testers, Anthropic reported that auto mode detected 89 per cent of harmful actions. By contrast, manual human review detected only 13.6 per cent of such actions. The company attributed this disparity to user habituation, noting that in previous iterations requiring manual oversight, testers approved 97 per cent of permission prompts, thereby reducing the efficacy of human intervention.
Claude Code Head Boris Cherny highlighted the team’s reliance on the new functionality in a post on X, stating that he and his colleagues have used auto mode exclusively for several months. Cherny remarked that he could not imagine returning to the previous permission prompt workflow, underscoring the internal confidence in the automated approach to balancing speed and control.
To mitigate risks associated with reduced human oversight, Anthropic is introducing additional safety mechanisms alongside the default setting. These include prompt injection screening and customizable hard deny rules designed to prevent data exfiltration. The company first unveiled a test version of auto mode in March 2026, positioning it as a method to streamline development processes while maintaining necessary safeguards.
The rollout follows a period of testing and refinement, with the August 14 date marking the full deployment for paid account holders. While the broader artificial intelligence landscape continues to evolve with developments such as OpenAI’s recent release of GPT-Live, Anthropic’s move specifically targets the security and efficiency of coding workflows. The company maintains that the automated system offers a more robust defence against harmful actions than traditional manual review methods.
