Event details from: Business Analytics Seminar: Behnam Tajfar (GenAI specialist ING)

Date
13 November 2025
Time
12:00 – 13:00

Abstract:

This work introduces a cutting-edge paradigm in AI safety and red-teaming: an autonomous, hierarchical Multi-Agent Self-Evolving (MASE) system designed to rigorously stress-test and continuously uncover vulnerabilities in LLM-based guardrails.

The architecture leverages a multi-tiered intelligence framework centered on meta-learning. At the foundational level, a Generator Agent produces adversarial linguistic inputs, guided by tactical instructions from a Strategist Agent. The Strategist dynamically adapts its tactics by evaluating performance metrics such as Verified Bypass Ratio. Above this layer, the Meta-Strategist Agent delivers the true innovation—performing high-fidelity introspection across extended epochs, critiquing and refining the system’s decision-making strategies to enable continuous improvement.

This hierarchical self-evolution allows the system to transcend fixed optimization strategies, discovering novel paradigms such as exploiting subtle keyword hierarchies within guardrail logic. A rigorous verification pipeline ensures reliable feedback loops, fostering a robust internal model of both successful attack vectors and defense mechanisms. Ultimately, this approach marks a significant step toward recursively self-improving AI capable of autonomously probing and challenging the boundaries of AI safety systems, offering critical insights for building resilient and trustworthy agentic AI and future AGI systems.

General information:

This seminar will be organised in a hybrid setup. If you are interested in joining this seminar, please send an email to the secretariat of Amsterdam Business School at secbs-abs@uva.nl.