Claude Code Auto Mode: Balancing Safety and Efficiency
Claude Code auto mode: a safer way to skip permissions
AI BriefingAnthropic has introduced “Auto mode” for Claude Code to address “approval fatigue” caused by users frequently having to approve actions manually. The mode sits between fully manual approval and the dangerous option of skipping permissions altogether, using a two-layer defense system: at the input layer, a server-side prompt-injection detector scans tool outputs; at the output layer, a transcript classifier based on the Sonnet 4.6 model assesses the risk of an action before it is executed. The classifier uses an efficient two-stage design: it filters quickly first and invokes chain-of-thought reasoning only when necessary. Its goal is to block dangerous actions, such as overly aggressive behavior, accidental mistakes, and prompt injection, while allowing most safe actions to proceed without confirmation. Internal tests showed that users would otherwise have approved about 93% of manual prompts.
Source: Anthropic: Engineering · anthropic.com