Session

Move Fast, Stay Secure: Enterprise AI Agent Security in Practice

Overview

ExperienceIn Person
TrackGovernance & Security
IndustryEnterprise Technology
TechnologiesUnity Catalog, Databricks Agents
Skill LevelIntermediate
DOWNLOAD SESSION SLIDES

Agentic AI is moving from POCs to mission-critical workloads- and the attack surface is expanding fast. As LLM-driven agents reason, plan, and act across enterprise data and tools, a new class of risk emerges: the "Lethal Trifecta," where sensitive data access, untrusted inputs, and external actions converge to make prompt injection attacks exploitable end-to-end.

 

Karol Piekarski (Experian) and Nishith Sinha (Databricks) present the Databricks Agentic AI Security Framework (DASF v3), drawing on both Databricks' own AI infrastructure which handles millions of daily queries, and patterns from Fortune 100 deployments. The session walks through how the Databricks platform breaks the Lethal Trifecta at each layer: constraining what agents can access, hardening how they process inputs, and limiting what a compromised agent can actually do.

 

Finally, see how the Databricks Security team built, open-sourced, and actively uses BlackIce, a containerized AI red teaming toolkit, to shift security from a point-in-time exercise to continuous, automated validation. Leave with a practical framework for moving fast on AI without compromising governance or security.

Session Speakers

Karol Piekarski

/Lead DevOps Engineer
Experian

Nishith Sinha

/Senior Engineering Manager- AI Security
Databricks, Inc.

Full Summary

Securing agentic AI: inside the lethal trifecta and how to defend against it

Agentic AI shifts security from a model-centric question to a system-wide discipline. When autonomous planning, tool use, MCP integrations, and memory are wired together, a single prompt injection can trigger data exfiltration or misuse of capabilities. The session outlines how an updated Databricks AI Security Framework, the lethal trifecta model, and continuous red teaming combine to harden production agents, drawing on security best practices for the Databricks Platform. Teams building on the platform can also consult the Databricks trust and security features best practices for a broader set of hardening recommendations that apply across the full Databricks Platform.

FAQ


A jailbreak tries to make a model violate its own policies. A prompt injection manipulates an agent through untrusted input so it follows attacker instructions, often to exfiltrate data or misuse tools. Agentic systems amplify prompt injection because model outputs trigger real actions.

Shared service principals collapse many users into one identity and often become over-privileged. On-behalf-of authentication inherits the caller's Unity Catalog permissions, preserving least privilege and keeping data scoped to each user.

Yes. Guardrails and LLM-as-judge checks are probabilistic and can be bypassed. Serverless egress controls and storage allow lists are deterministic, so even if an injection succeeds, the agent cannot reach attacker endpoints.

Continuously. Any change to prompts, MCP servers, skills, or dependencies can reopen vulnerabilities, so BlackEyes-style scans belong in CI/CD and should be run on demand whenever the agent's bill of materials changes.

Use care. BlackEyes performs real adversarial testing that exercises hosting infrastructure. Confirm contractual permission with model and infrastructure providers, and scope tests appropriately since deeper scans can take longer than normal deployments.