Key takeaways
- How to use a crawl, walk, run framework to pick generative AI use cases: look for high volume, high reward, low risk, and high-provenance data before you start.
- Why starting with a small, limited-blast-radius use case (not the flashiest problem) builds trust with mission owners before you tackle harder problems.
- The case for sovereign AI: LLM-agnostic, air-gapped, or on-prem deployments (H2O.ai noted it can host open-source models from Hugging Face, roughly 50 of them, fully on prem) that keep data inside the wire and let agencies interrogate models for drift, bias, and hallucination.
- How privileged access management risk changes with generative AI, including an example of using generative AI to monitor privileged session recordings for anomalies, discussed in the context of lessons learned from the SolarWinds breach.
- Why workforce upskilling and keeping a human in the loop (and on the loop) matter as much as the technology, including the case for third-party assessment of models separate from whoever built them.
- A real example of generative AI resolving conflicting federal guidance: inference cut seven agencies' twelve different answers for how long to boil water before drinking down to three, based on altitude.
Overview
In this webinar, H2O.ai and Carahsoft host federal cybersecurity leaders, including former USCIS CISO Shane Barney, for a conversation on how government agencies are unlocking value with generative and agentic AI. They walk through decision frameworks for AI adoption, picking low-risk, high-reward use cases, and the case for sovereign, air-gapped AI deployments that keep sensitive data inside the wire. The discussion also covers new privileged access risks tied to AI, workforce upskilling, and real examples like resolving conflicting federal water-safety guidance and lessons from the SolarWinds breach.