Incident Triage Drill
Work through a realistic on-call scenario for AI and cloud security incidents. Pick the incident type you want to drill.
30 scenarios across 8 incident types
What this is. Each drill is an incident scenario: a system context, a SIEM alert that fired, and a sequence of triage decision steps. You click through the decisions, see which calls were right, and get a full debrief showing what thorough coverage of that incident type looks like. The debrief downloads as a Markdown file you can keep.
Who it is for. Security analysts, cloud engineers, and defenders who work with AI systems and want to drill the triage patterns for AI-specific and cloud-native attacks before they encounter them in production.
Incident types
- Prompt Injection5 scenarios
Instructions smuggled into model input through a user field, a retrieved document, or a tool result, overriding what the system prompt told the model to do.
- Inference API Abuse3 scenarios
A model endpoint driven outside its intended use: bulk querying to extract behavior, cost abuse, or jailbreak attempts at volume.
- Training Data Poisoning4 scenarios
Corrupted examples introduced into a training or fine-tuning set, so the resulting model carries the attacker’s behavior as a property of its weights.
- Model Exfiltration3 scenarios
Weights, adapters, or a functional copy of a model leaving the environment, either downloaded directly or distilled out through the API.
- RAG Pipeline Poisoning3 scenarios
Attacker-controlled content planted in a vector store or in a source the retriever trusts, so the model repeats it back as grounded fact.
- Agent Privilege Escalation6 scenarios
An agent chaining its own tool access into permissions no single tool was meant to grant, usually without any individual call looking abnormal.
- IAM Compromise (AI Workloads)3 scenarios
Stolen or over-broad credentials on the identities AI workloads run as: service accounts, assumed roles, and key material held by inference infrastructure.
- ML Supply Chain Attack3 scenarios
A compromised model artifact, dataset, or ML dependency pulled in from a registry or a model hub and trusted by everything downstream of it.