AI Security Wiki

A versioned map of the Databricks AI Security Framework, compiled for people and agents. Start an audit · Português · MCP · Sources

Snapshot 2026-10-22, DASF 3.0: 97 risks, 73 controls, 13 components, 7 deployment patterns. These counts describe the compendium, not all possible AI security knowledge.

Navigate by architecture

• 1. Raw data — 11 risks
• 2. Data preparation — 4 risks
• 3. Datasets — 3 risks
• 4. Data catalog governance — 2 risks
• 5. Machine learning algorithms — 4 risks
• 6. Evaluation — 3 risks
• 7. Machine learning models — 4 risks
• 8. Model management — 4 risks
• 9. Serving: inference requests — 13 risks
• 10. Serving: inference responses — 6 risks
• 11. Operations — 1 risks
• 12. Data and AI platform — 7 risks
• 13. Agentic AI — 35 risks

Select every deployment pattern

• Predictive ML Models
• RAG - LLMs
• Fine-tuned LLMs
• Pre-trained LLMs
• Foundational LLMs
• External Models
• Agentic AI

Investigation and learning

• Audit method
• Security research: PentAGI, Cloudflare and Red Blue Field
• Evidence-driven audit playbook
• Jason Haddix study: recon, tools and reports
• Map attack surface skill
• Pocket cybersecurity specialist
• Jason Haddix posts and media
• Evaluate models, guardrails and applications with spikee
• Agent and MCP trust boundaries
• Glossary
• Finding template
• Source-listed third-party tools
• Source inconsistencies
• All controls
• Source and update policy
• Change log