FinSec: A Generative Defensive Agent for Financial Dialogue Security via Multi-Stage Rollout

Abstract With the rapid adoption of large language models(LLMs) and agents in financial services, ensuring conversational safety under stringent regulatory constraints has become increasingly challenging. However, existing methods typically rely on simple semantic or fixed-rule sets, which are inadequate for dynamic, multi-turn interactions and for specific financial compliance requirements. To address these issues, we propose FinSec, a generative defensive agent with a multi-stage rollout mechanism for financial-agent dialogues. FinSec integrates suspicious behavior pattern detection, delayed-risk and adversarial inference, and semantic safety analysis to risk fusion decision-making, and ultimately to the autonomous generation of structured defensive intervention strategies. This enables structured, interpretable, end-to-end identification of real-world financial risks. While preserving model utility, FinSec significantly improves robustness in detecting high-risk conversations. Experimental results show that FinSec achieves state-of-the-art performance. For overall detection, FinSec attains an F1 score of 90.13%, outperforming baseline models by 6–14%; its attack success rate (ASR) is reduced to 9.09%, significantly limiting unsafe outputs. Furthermore, FinSec achieves about 9.7%improvement in the area under the precision–recall curve (AUPRC) over generic frameworks. Moreover, in the evaluation of the utility–risk trade-off, FinSec reaches a composite score of 0.909, providing strong and efficient safety guarantees for financial-agent conversations.