SecureClaw preprint reports gains in blocking agent attacks and data leaks

Researchers presenting SecureClaw report that their architecture limits unauthorized actions and sensitive-data exposure by tool-using AI agents. In their benchmark evaluation, it recorded 0% attack success on Agent Security Bench, 0.64% on AgentDojo, and 3.23% overall leakage on AgentLeak's attacked parity lane while retaining usable task utility. A trusted gateway replaces protected values with reference IDs and releases selected information through short planning summaries. A separate trusted executor performs only the exact external request authorized by policy, using a preview-and-commit process. The language model cannot directly retrieve secrets from the IDs or execute external changes. The preprint's leakage result covers both final outputs and internal relays; the architecture still deliberately exposes selected information through its summaries.