Episode 11Operations

AI Incident Response — What’s Different

Bridge Guide + Threat Model

Video coming soon

Security teams have mature IR processes for traditional systems. But AI-specific incidents require new procedures: model quarantine, training data breach investigation, prompt injection containment — these are IR procedures that don’t exist in most security programs.

“When an AI system is compromised, your traditional incident response playbook won’t cover it. Model quarantine, training data breach investigation, prompt injection exploitation containment — these are IR procedures that don’t exist in most security programs. I built them.”

Architecture Diagrams

AI IR decision tree (detection → classification → containment → investigation → recovery)

Model-level vs. infrastructure-level incident classification

Containment options matrix by incident type

Build Notes

IR decision tree: detection → classification (model-level vs. infrastructure-level) → containment → investigation → recovery
Model-level containment: rate limit API, enable output filtering, switch to known-good version
Infrastructure-level: network isolation plus model registry lockdown
AI incidents require coordination between security, ML engineering, and legal teams

Lessons Learned

Run tabletop exercises for AI-specific scenarios before the real incident happens
Model versioning is your rollback mechanism — if you don’t version, you can’t roll back
The hardest IR decision is whether to take a production model offline — build the criteria in advance
AI IR playbooks should be tested quarterly, just like network IR playbooks

Discussion

Has your organization ever run a tabletop exercise for an AI security incident? If you were to run one tomorrow, what scenario would you choose — prompt injection, model extraction, or training data poisoning?