Webinar

September 1, 2025

Jailbreaking LLMs and Agentic Systems

Cisco Senior Director of AI Research Amin Karbasi joins Virtue AI Co-founder Sanmi Koyejo for a conversational session exploring cutting-edge research on jailbreak attacks and defenses, Evaluation methods shaping the future of AI safety and How builders can design systems to be secure by default

Summary:
This webinar explores how LLMs and agentic AI systems can be systematically jailbroken through automated, scalable attacks, and how enterprises can respond with layered, continuously evolving defense strategies.

Key points:

  • Jailbreaking LLMs is often a search problem, enabling highly effective automated attack methods like Tree of Attacks (TAP) and adversarial reasoning.
  • Modern attacks increasingly use multi-turn and decomposition strategies that hide malicious intent across benign-looking queries.
  • Automated red teaming can outperform human attackers in scale, cost, and consistency.
  • Enterprise risk is best understood through business impact (cost, data exposure) rather than raw attack success rate.
  • Traditional cybersecurity is insufficient due to AI systems being dynamic and context-dependent, requiring continuous monitoring.
  • Effective defense relies on a “Swiss cheese” model: stacking alignment, prompt defenses, reasoning filters, guardrails, and monitoring.
  • Agentic systems expand risk from bad outputs to harmful actions, increasing the stakes significantly.
  • Enterprises should focus on safe tool access, employee education, and endpoint monitoring to reduce exposure.

Strengthen Your AI Posture Today

Virtue AI brings control, governance, and resilience to enterprise AI.