AI Landscape

• Map foundation models and applications
• Trace the data supply chain
• Account for plugins and integrations
• Note the cloud and edge infrastructure

Model Manipulation

• Guard against poisoning and backdoors
• Watch for model inversion of sensitive data
• Verify behavior has not been altered

Prompt Injection

• Treat user input as a weapon
• Expect attempts to bypass guardrails
• Block unauthorized actions from prompts

Data Exfiltration

• Watch for subtle leakage in outputs
• Check embeddings and side channels
• Limit what the model can reveal

Agent Risks

• Assume agents can be coerced
• Guard against misdirection and repurposing
• Constrain what autonomous agents may do

Adversarial ML

• Expect evasion attacks on vision and audio
• Test decision systems for manipulation
• Validate inputs before trusting outputs

Exposure Paths

• Trace input through model to output
• Account for logging and storage
• Break the chain where you can

Exposure Points

• Strip device ID, metadata, and location
• Limit biometrics and file uploads
• Control prompt content you submit