720 Attacks, 0 Successes: Claude Code Defaults to Auto-Approval, AI Clicks 'Agree' for You
The era of manually approving every action for Claude Code is ending. Starting August 14th, Claude Code will default to an "Auto Mode" for Pro, Max, and Team plans, where AI automatically approves actions instead of prompting users for permission each time. This change is driven by a security report from Trajectory Labs commissioned by Anthropic, which tested 720 attack scenarios across Claude's latest models with zero successful breaches.
The defense relies on a three-layer system: the aligned model itself, an input-side probe to detect hijacking attempts, and an output-side classifier to vet actions before execution. Anthropic claims this stack has made prompt injection attacks undemonstrable even internally. Tests showed the Auto Mode blocked 89% of dangerous commands, far surpassing the 13.6% interception rate by human users, who often develop "approval fatigue."
However, experts like Simon Willison express caution, warning of potential blind spots. Independent research suggests vulnerabilities may persist, particularly with indirect attacks like malicious commands in third-party packages. While Auto Mode improves efficiency and reduces human error from repetitive prompts, it shifts approval authority from the user to AI. The article concludes that while the system is more reliable than fatigued users, it does not eliminate risk, and ultimate responsibility remains with the human user.
marsbitHace 44 min(s)
