MAGE: Shadow Memory Defense Reduces Agent Tool-Chaining Attacks from 100% to 8.3% Success
arXiv·medium signal
MAGE introduces a safety-focused 'shadow memory' inspired by shadow stacks in systems security. It maintains a dedicated memory that distills safety-critical context across the agent's full execution trajectory, evaluating pending actions before execution. The approach is the first framework to detect and mitigate long-horizon multi-step threats using agentic memory, with negligible performance overhead.