[Live Session] Choosing safer LLMs: From LLM benchmarks to your production agents

Picking an LLM by leaderboard rank is a gamble. A model can top the charts and still fail on the risks that matter for your use case — jailbreaks, hallucinations, or biased outputs. Latest results from Phare, our independent LLM benchmark that evaluates 71 models on safety and security, show the gap between providers widening.

In this live session, Giskard AI researcher Pierre Le Jeune will walk through our research projects and turn the results into a practical method for selecting and monitoring the models behind your AI agents.

Agenda:

• Safety & security gaps across model providers: Results from all 71 models now on Phare, including 13 new entrants (GPT 5.5, Claude 5 Sonnet, Kimi K2.6, DeepSeek V4, and others).
• Discovering stereotypes in open-ended generation: The gap between what models recognize as harmful and what they generate, and why stereotypes shift with the prompt language.
• From benchmark to your production agent: Practical recommendations from both studies, and how to go from public benchmarks to red teaming your own agents.
• Conclusion and Q&A

Who Should Attend: CISOs, security architects, Heads of AI, AI/ML engineers, AI Product Managers, and researchers who select, deploy, or audit LLMs in production.