Research from Dreadnode and co-authors of Microsoft's PyRIT describes an AI agent that autonomously selected attack strategies, applied prompt transforms, and executed 674 adversarial attacks against Meta's Llama Scout (17B) in roughly three hours, achieving an 85% overall attack success rate and 100% on several techniques. The work illustrates a structural shift: agent orchestration is collapsing the operational overhead of LLM red teaming, making continuous adversarial assessment feasible for a single operator rather than a specialist team. Critically, the same accessibility gains apply to malicious actors, and the paper was published without coordinated disclosure with Meta. For financial institutions deploying or procuring LLM-based systems, this signals that the cost and frequency assumptions underlying current AI security testing programs and vendor due-diligence frameworks are likely already outdated.
199 other tracked developments touching the same claims.