CrowdStrike Falcon AIDR with NVIDIA NeMo Guardrails Securing AI Agents

CrowdStrike Falcon AIDR has integrated NVIDIA NeMo Guardrails (v0.20.0), offering a robust defense for AI agents transitioning from development to production environments. The integration addresses critical risks associated with AI agents, such as unauthorized data access, compromised transactions, and compliance violations. NVIDIA NeMo Guardrails provides content safety, PII detection, jailbreak detection, and topic control, while Falcon AIDR manages data access, controls agent responses, and ensures policy compliance. This combined solution enables organizations to confidently deploy AI agents, mitigating risks associated with prompt injection attacks, sensitive data exposure, and malicious content execution. Falcon AIDR includes 75+ built-in classification rules and supports custom data classification.

Attack Chain

1. Initial Access via Prompt Injection: An attacker crafts a malicious prompt to inject unauthorized commands or queries into the AI agent via a user interface or API.
2. Bypass of Initial Guardrails: The prompt injection attempts to evade initial input validation and content filtering mechanisms.
3. Data Exfiltration: The compromised agent, under attacker control, is tricked into accessing sensitive data, such as customer PII, PHI, or internal code repositories.
4. Unauthorized Transaction Execution: The attacker manipulates the agent to execute unauthorized transactions, like fraudulent refunds or financial transfers.
5. Lateral Movement: The compromised agent leverages its access to internal tools and data sources to move laterally within the organization's network.
6. Code Injection: In software development environments, the attacker injects malicious code into the codebase via a compromised AI coding assistant.
7. Credential Access: The agent, tricked by the attacker, exposes hardcoded secrets or internal repository references, enabling further access.
8. Compliance Violation: The AI agent, now controlled by the attacker, violates compliance boundaries by exposing sensitive data or engaging in prohibited activities.

Impact

A successful attack on an AI agent can lead to significant consequences, including exposure of customer data, unauthorized financial transactions, and compliance violations. For example, in financial services, compromised agents could expose account numbers and SSNs, leading to financial fraud and regulatory penalties. Healthcare organizations could suffer breaches of PHI, risking patient privacy and legal repercussions. In software development, code injection can compromise software integrity and create supply chain vulnerabilities. The impact extends across sectors where AI agents manage sensitive data and processes.

Recommendation

• Deploy Falcon AIDR with NVIDIA NeMo Guardrails (v0.20.0) to provide runtime protection for AI agents.
• Create custom Falcon AIDR policies to tailor detection and response to specific security requirements.
• Utilize Falcon AIDR's API to apply policies at critical points in AI agent workflows, such as chat input sanitization and data ingestion.
• Monitor Falcon AIDR's detection logs for prompt injection attempts and other malicious activities targeting AI agents.
• Implement Falcon AIDR's redaction capabilities to prevent exposure of sensitive data like PII and PHI.
• Configure Falcon AIDR policies to block code injection attempts and detect hardcoded secrets within AI coding assistants.