security-drift

This paper audits self-evolution mechanisms in financial AI agents, revealing capability improvements alongside security risks such as prompt injection drift and execution-interface mismatches, emphasizing the need for holistic auditing.