Skip to main contentSkip to main content
    Back to Glossary
    security
    2 min read

    LLM Security

    LLM security addresses the unique risks of deploying Large Language Models, including prompt injection, data leakage, and adversarial attacks on AI systems.

    LLM security covers the specific vulnerabilities and risks associated with deploying Large Language Models in production.

    Top LLM security risks (OWASP LLM Top 10): 1. Prompt Injection: Manipulating model behavior via crafted inputs 2. Data Leakage: Model exposing training data or sensitive info 3. Inadequate Sandboxing: LLM with excessive system access 4. Supply Chain Vulnerabilities: Compromised models or libraries 5. Training Data Poisoning: Malicious data in training sets 6. Overreliance: Trusting LLM output without validation 7. Insecure Plugin Design: Vulnerable integrations 8. Excessive Agency: LLM making autonomous decisions

    Mitigations: - Input validation and output filtering - Principle of least privilege for LLM access - Human-in-the-loop for sensitive actions - Model output validation - Rate limiting and monitoring - Content filtering and guardrails

    Why It Matters

    As organizations rapidly adopt LLMs, new attack vectors emerge that traditional security tools cannot detect. Prompt injection can bypass safety guardrails, LLMs can leak sensitive training data, and AI agents with excessive permissions can be manipulated to take harmful actions. The OWASP LLM Top 10 provides a framework for addressing these risks, and organizations deploying AI must integrate LLM-specific security controls alongside traditional application security.

    Key Points

    Prompt injection is the SQL injection of AI
    LLMs should have minimal system permissions
    Never trust LLM output without validation
    Monitor for unusual patterns and abuse
    Consider data leakage in training and inference

    Applicable Compliance Frameworks

    Related Terms

    Frequently Asked Questions

    What is prompt injection?

    An attack where malicious input causes the LLM to ignore instructions or perform unintended actions. Similar concept to SQL injection but for AI.

    How do I prevent data leakage from LLMs?

    Don't put sensitive data in prompts, use access controls, implement output filtering, and consider private/on-premise models for sensitive use cases.

    Need Help with LLM Security?

    Our experts can help you understand and implement the right controls for your organization.