New whitepaper outlines the taxonomy of failure modes in AI agents