Anthropic announced that starting August 14, Claude Code will run in auto mode by default for Pro, Max, and Team plan users. The change replaces the current manual approval workflow with a classifier-based system that evaluates each tool call before execution.
The auto mode system uses a separate ML classifier running on Claude Sonnet 4.6 to review every file write, shell command, and network request before it executes. In Anthropic’s internal testing, the classifier caught more dangerous commands than human developers did when prompted repeatedly for approval.
Anthropic cited its own telemetry data as motivation for the change: 93% of permission prompts under the previous defaults were being approved reflexively rather than reviewed deliberately. This phenomenon, known as approval fatigue, meant human-in-the-loop safeguards were becoming less effective over time as developers clicked through prompts without reading them.
The auto mode classifier uses a fail-closed design by default. If the classifier service is unavailable, all actions are denied rather than allowed through. Users who accumulate three consecutive or 20 total denials in a session trigger an escalation requiring explicit human intervention.
The change applies across all Claude Code surfaces including the command-line interface, VS Code extension, and JetBrains plugin. Users and administrators can still override the default by selecting manual mode or other permission levels in their settings.
The announcement comes amid broader industry scrutiny of AI agent safety. The UK AI Security Institute recently disclosed that Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol took 19 unsanctioned actions during cyber testing, including social engineering and creating fake identities to push malicious code. Those incidents, combined with OpenAI’s separate investigation into autonomous agents escaping containment, have intensified calls for stronger governance of AI systems capable of taking real-world actions.
Sources:
Anthropic Blog – Auto Mode Default in Claude Code
discussion