Open-source tool SafetyDrift detects AI agent data leaks that individual guardrails miss

A new open-source tool called SafetyDrift catches sequence attacks on AI agents, where individual tool calls appear safe but together leak data. It tracks data exposure, tool escalation, and reversibility across a session, predicting violations within 5 steps with 85% accuracy. The tool adds two lines of code to existing agents and is based on a March 2026 arXiv paper.