Chatbot Security and Privacy Flashcards

What is adversarial input in chatbot security?

Answer: Specially crafted user messages designed to cause the chatbot to malfunction, reveal sensitive data, or produce harmful outputs

Adversarial inputs are deliberately engineered to exploit weaknesses in a chatbot's model or logic, often to bypass filters, extract hidden information, or generate unsafe content.

What is session hijacking in the context of a chatbot application?

Answer: An attacker stealing a valid session token to impersonate an authenticated user and access their chatbot conversation

Session hijacking occurs when an attacker obtains a user's session token (via XSS, network sniffing, etc.) and uses it to impersonate that user in subsequent chatbot interactions.

How does input sanitization protect a chatbot from security threats?

Answer: It strips or escapes dangerous characters and patterns from user input before the chatbot or backend systems process it

Input sanitization removes or neutralizes characters and sequences (like SQL operators or script tags) that could be exploited in injection or XSS attacks before they reach backend systems.

What role does audit logging play in chatbot security?

Answer: It creates a tamper-evident record of all chatbot interactions and system events, enabling forensic analysis after a security incident

Audit logs provide an immutable record of who did what and when, which is essential for investigating breaches, meeting compliance requirements, and detecting anomalous behavior.

Which US state regulation specifically governs data privacy rights for residents using chatbots and online services?

Answer: CCPA (California Consumer Privacy Act)

CCPA grants California residents rights over their personal data collected by businesses, including chatbots — such as the right to know, delete, and opt out of data sale.

What is the primary risk of training a chatbot on raw, unsanitized customer support data?

Answer: The chatbot may learn to reproduce or leak sensitive PII from training examples during conversations

If training data contains PII such as names, account numbers, or medical details, the model can memorize and inadvertently reproduce this information in responses to unrelated users.

What best practice ensures chatbot users in a multi-tenant SaaS environment cannot access each other's conversation data?

Answer: Strict tenant isolation enforced at the data layer, ensuring each tenant's data is logically or physically separated

True tenant isolation at the data layer (separate schemas, databases, or row-level security policies) prevents one tenant's queries from ever accessing another tenant's records, even in case of application bugs.