Google SRE Prodcast The One with Parker Barnes, Felipe Tiengo Ferreira, and AI
Feb 5, 2026
Parker Barnes, product manager for Gemini model-level safety, focuses on policies and deployment safety. Felipe Tiengo Ferreira, tech lead on Gemini safety, handles monitoring and rapid response. They talk about model safety frameworks and policy design. They cover prelaunch testing, automated red teaming, layered defenses like system instructions and classifiers. They discuss drift detection, observability, and the pressure of rapid AI development.
AI Snips
Chapters
Transcript
Episode notes
Safety Includes Frontier Risks
- Safety spans content policy violations and broader frontier risks like cybersecurity and CBRN.
- Parker Barnes highlights a formal Frontier Safety Framework to address those systemic harms.
Safety Is Squishy And Contextual
- Model interactions are open-ended so safety becomes subjective and continuous.
- Felipe Tiengo Ferreira notes safety and quality blur in generative systems.
Use Automated Red Teaming Prelaunch
- Run Automated Red Teaming (ART) to explore adversarial and real-user inputs before launch.
- Use ART to catch egregious failures across the infinite input/output space.