Secure Your AI Lifecycle with AI Protection in Security Command Center

Key Concepts:

• AI Security Dashboard
• AI Lifecycle Security
• AI Inventory (Projects, Foundation Models, Custom Models, Datasets)
• Sensitive Data
• Risk Engine (Virtual Red Teaming)
• Data Poisoning
• Model Armor (Guardrails)
• Prompt Injection, Jailbreak, Offensive Content Generation
• Responsible AI Filters
• Vertex AI Integration
• LLM Hijacking
• Coordinated Prompt Injection Attack
• SecOps

1. AI Security Dashboard and Visibility:

• The AI Security Dashboard within Security Command Center provides a single pane of glass for securing the entire AI lifecycle.
• It starts with visibility into AI inventory:
  • 14 projects with AI usage.
  • 18 instances of foundation models (e.g., Llama, Gemini) used for inference.
  • 21 custom models tuned/trained using data from 27 datasets.
  • 18 datasets containing sensitive data.

2. Risk Engine and Prioritization:

• The AI inventory is fed into a risk engine.
• The risk engine performs virtual red teaming to identify toxic combinations and attack paths that could compromise sensitive AI resources.
• The riskiest AI issues are prioritized in a list.
• Example: Data poisoning risk where a publicly exposed bucket contains sensitive model tuning data. Compromising this data could impact the model's behavior.

3. Data Poisoning Remediation Example:

• The demo focuses on a data poisoning risk involving a Gemini model.
• A detailed lineage view shows the connection between:
  • Sensitive data categories in the tuning data.
  • A publicly accessible bucket storing the tuning data.
  • The Gemini model instance deployed through an endpoint.
• Remediation involves narrowing the permission set on the publicly accessible bucket to mitigate the data poisoning risk.
• Specific remediation guidance is provided for the impacted bucket.

4. Runtime Risk and Model Armor:

• The demo shifts to runtime risks, specifically focusing on model inference.
• 4 instances of foundation models used for inference are not protected by Model Armor.
• Lack of Model Armor exposes the organization to risks like:
  • Brand reputation damage.
  • Data loss.
  • Prompt injection.
  • Jailbreak attacks.
  • Offensive content generation.

5. Model Armor Configuration and Vertex AI Integration:

• The demo shows how to enable Model Armor for a Gemini 2.5 Pro instance.
• The configuration flow allows selecting different detectors:
  • Malicious URL detection.
  • Prompt injection detection.
  • Jailbreak detection.
  • Sensitive data protection (data loss prevention).
  • Responsible AI filters.
• Model Armor integrates with Vertex AI, allowing guardrails to be applied without application changes.
• Enabling the Vertex AI integration automatically routes prompts and responses to Model Armor.
• This provides baseline visibility into potential issues without requiring developer intervention.

6. Model Armor Insights and Threat Detection:

• Projects protected by Model Armor feed insights into the AI Security Dashboard.
• The dashboard displays the distribution of harmful vs. not harmful content.
• It also shows the categories of violations detected by Model Armor.
• The system detects active threats against AI, such as:
  • Potential LLM hijacking (anomalous model usage).
  • Potential coordinated prompt injection attacks.
• Users can investigate these threats and initiate response workflows through integrated SecOps capabilities.

7. Conclusion:

• The demo showcases how to:
  • Automate the discovery of AI inventory.
  • Use a risk engine to prioritize AI security issues.
  • Detect active threats against AI.
  • Secure the AI lifecycle from training to inference.