Free AAISM sample questions

Correct answer: Secure Multi-Party Computation (SMPC) or Differential Privacy noise addition

In Federated Learning, sending raw gradients to a server can still leak information about the training data (model inversion/inference attacks). To prevent this at the aggregation step, Secure Aggregation (often using SMPC) or adding Differential Privacy (noise) to the updates is required to ensure the server sees only the aggregate update, not individual contributions. sequenceDiagram participant Client1 participant Client2 participant Server Client1->>Client1: Train Local Model Client2->>Client2: Train Local Model Client1->>Server: Send Gradient (Encrypted/Noisy) Client2->>Server: Send Gradient (Encrypted/Noisy) Note over Server: Secure Aggregation Required Here Server->>Server: Update Global Model Server-->>Client1: Send New Global Model Server-->>Client2: Send New Global Model

Correct answer: Prohibit the input of Personally Identifiable Information (PII) and Intellectual Property (IP) into public AI prompts

The most critical risk with public GenAI tools is the training of the provider's models on user data. An AUP must explicitly prohibit entering sensitive data (PII, IP, secrets) into public tools, while allowing use for non-sensitive tasks to balance security and productivity.

Correct answers: AI systems used for credit scoring and evaluating creditworthiness; AI systems used for recruitment and employee selection

AI systems evaluating creditworthiness of natural persons or establishing their credit score are High Risk as they determine access to essential private services.

Under the EU AI Act, systems used in employment, workers management, and access to self-employment (e.g., CV screening software) are classified as High Risk due to their potential impact on fundamental rights.

Correct answer: Evasion Attack

This describes an Evasion Attack (often using adversarial examples). The attacker manipulates the input data (inference time) with perturbations (noise) to cause the model to make a prediction error. It does not tamper with the training data itself.

Correct answer: Model Card

A Model Card (introduced by Google researchers and widely adopted) is a standardized document that details performance characteristics, intended use, limitations, and ethical considerations of an AI model.

Correct answer: Requiring a cryptographic signature for model artifacts and validating it against a trusted root of trust before loading

Cryptographic signing of model artifacts (e.g., using Sigstore or similar tools) ensures integrity and authenticity. Validating the signature before loading prevents the execution of tampered models or models from untrusted sources, directly addressing the risk of supply chain injection attacks.

Correct answer: False

False. Explainability is critical for security forensics. If an AI system behaves maliciously (e.g., due to an adversarial attack or poisoning), security analysts need explainability tools (like SHAP or LIME) to understand *why* the model made that decision and identify the root cause of the breach.

Correct answer: Prompt Injection (Jailbreaking)

Prompt Injection (specifically direct injection or jailbreaking) involves manipulating the input to override the model's original instructions or safety constraints. This allows the attacker to control the model's output.

Correct answer: Analyzing CASB (Cloud Access Security Broker) and web gateway logs for traffic to known AI domains

Analyzing network traffic via CASB or Secure Web Gateways (SWG) is the most reliable technical method. It identifies actual traffic to known AI service domains (e.g., openai.com, huggingface.co) regardless of whether employees have reported them.

Correct answer: Confidential Computing (Trusted Execution Environments / Enclaves)

Confidential Computing uses hardware-based Trusted Execution Environments (TEEs) or enclaves (like Intel SGX, AWS Nitro Enclaves) to isolate data in memory. This ensures that even the cloud provider's hypervisor or administrators cannot view the data while it is being processed (encryption-in-use).