A preprint from Queen's University and University of Waterloo taxonomizes generative AI security and safety threats across five escalating levels, from direct prompt injection through multi-agent cross-organizational compromise, finding that defensive responses and governance arrangements consistently lag behind capability and attack-surface expansion. The paper documents that human detection of AI-generated images is near chance (62%), frontier inference costs have fallen to sub-dollar per million tokens, and agentic deployments now expose file systems, APIs, and tool chains to adversarial manipulation with limited institutional safeguards in place. For financial institutions, the convergence of low-cost targeted impersonation, weakened voice and identity authentication, and expanding agentic attack surfaces represents a structurally growing threat to fraud controls, vendor-risk frameworks, and enterprise AI deployment governance.
434 other tracked developments touching the same claims.