Data poisoning is an adversarial machine learning technique in which attackers deliberately manipulate information used during model training or adaptation to influence subsequent model behavior. Option C is correct because the attacker contaminates training records with misleading examples that incorrectly label insecure programming practices as acceptable.
During model development, training examples contribute to the patterns learned by the model.
If malicious records are incorporated into the training dataset, the model may internalize incorrect associations and reproduce them when generating recommendations.
For example, corrupted training examples might repeatedly describe unsafe authentication practices or inadequate input validation as recommended implementations.
A coding assistant influenced by these examples could generate insecure code despite appearing to follow legitimate programming instructions.
Option A is incorrect because prompt injection typically targets model behavior during inference through adversarial instructions.
Option B is incorrect because model inversion attempts to recover sensitive training information or characteristics through model interactions.
Option D is incorrect because overfitting occurs when a model learns training-specific patterns that fail to generalize effectively. It does not necessarily involve malicious manipulation.
Defensive measures include training-data provenance verification, integrity controls, dataset auditing, anomaly detection, and security-focused model evaluation.
Specific Cisco AITECH Course Guide References:
Generative AI Security and Privacy Fundamentals
Debugging and Correcting AI-Generated Outputs
Exam Objective 3.3: AI-specific security threats and risks
Exam Objective 3.4: AI governance and risk management
===================
Submit