Adversarial suffix attacks and multilingual input manipulation are commonly used to test whether an AI chatbot can be tricked into bypassing safeguards or executing unintended behavior. The primary goal is to evaluate the chatbot's resilience against prompt injection and adversarial manipulation techniques.

Third-party pre-trained models often rely on external libraries and components that may include unvalidated or malicious dependencies. Risks such as typosquatting can introduce compromised packages, creating a direct supply chain security threat that can lead to system compromise or data exposure.

AAISM identifies training as the phase with the highest inherent risk because this is where:
- data poisoning can occur
- sensitive data may be exposed
- bias can be introduced
- model inversion risks originate
- security and privacy vulnerabilities are embedded

The data Preparation phase-covering sourcing, collection, labeling, cleansing, and provenance-presents the greatest inherent risk because it is where privacy, consent, representativeness, bias, quality, lineage, and legality must be established. Decisions and defects here propagate into training and downstream use, amplifying ethical, regulatory, and accuracy risks.
While Training can introduce additional risks (e.g., poisoning, leakage), these are frequently mitigated by controls that depend on having trustworthy prepared data. Monitoring and Maintenance are later-life-cycle phases oriented toward detection and correction; they are critical but inherently rely on the foundation set during Preparation.
References: AI Security Management™ (AAISM) Body of Knowledge: "AI Data Lifecycle Risks-Sourcing, Consent, Provenance, and Bias," "Risk Treatment Priorities Across the Lifecycle"; AAISM Study Guide:
"Data Preparation Controls and Quality Gates," "Bias and Privacy Risk Controls at Ingestion."

AAISM defines automation bias as the tendency of individuals to over-rely on AI-generated outputs even when contradictory real-world evidence is available. In this scenario, the driver ignores traffic signs and follows the AI's instructions, showing blind reliance on automation.
Selection bias relates to data sampling, reporting bias refers to misrepresentation of results, and confirmation bias involves interpreting information to fit pre-existing beliefs. The most accurate description is automation bias.