Privacy, Security, and AI Safety

Protecting Data Privacy

AI systems often process personal information, including names, financial records, location data, and health information. Poor data handling can expose individuals to privacy risks.

Important practices include:

• Collect only the information necessary for the intended purpose.
• Obtain appropriate consent or another valid legal basis when required.
• Restrict access to sensitive datasets.
• Encrypt data where appropriate.
• Define retention periods and delete information when it is no longer needed.
• Use anonymization or pseudonymization techniques when suitable, while recognizing that re-identification may still be possible.

Developers should also avoid placing confidential information into AI tools unless their data-handling policies and security controls permit it.

Securing AI Applications

AI applications face conventional cybersecurity risks as well as risks specific to machine learning and generative AI.

Examples include:

• Data poisoning: Manipulating training data to influence model behavior.
• Adversarial inputs: Carefully designed inputs that cause incorrect predictions.
• Prompt injection: Instructions hidden in user input or retrieved content that attempt to override an AI application's intended behavior.
• Sensitive data leakage: A model or application reveals confidential information.

Improving AI Safety

Security and safety controls should be designed into the application rather than added only after deployment.

Useful safeguards include input validation, access controls, output filtering where appropriate, secure data retrieval, rate limiting, testing against misuse, and human review for high-impact decisions.

For example, a customer-support chatbot should retrieve information only from sources the user is authorized to access. It should also provide a clear escalation path when it cannot answer safely or reliably.