Updating the taxonomy of failure modes in agentic AI systems: What a year of red teaming taught us