Code Story: Insights from Startup Tech Leaders The Gene Simmons of Data Protection - AI Inference-time Guardrails
Feb 11, 2026
Ave Gatton, Director of Generative AI at Protegrity and former atomic and optical physicist, discusses inference-time AI risks. She outlines prompt injection, data exfiltration, and agent misuse. She compares training vs inference threats. She walks through industry differences and practical secure-by-design guardrails for real-time protection.
AI Snips
Chapters
Transcript
Episode notes
Agents Have No Built-In Security Model
- Agents bridge users and external systems, creating unique security risks when given data access and communication abilities.
- Ave Gatton warns agents lack an internal security model and can be directed to exfiltrate or manipulate data if seized.
Training Risks Versus Inference Risks
- Training-time risks like model poisoning embed backdoors in model behavior via malicious fine-tuning data.
- Inference-time risks are tactical and immediate, where prompts or interactions cause malicious actions in the moment.
Design For Inevitable Jailbreaks
- Assume any deployed agent will be jailbroken at scale and design defenses accordingly.
- Focus on limiting the worst-case actions an agent can take rather than assuming perfect immunity.