Ethics and Security

Introduction

As generative AI moves from novelty to core business infrastructure, the risks shift from "does it work" to "can it be trusted, and can it be abused." This topic covers responsible AI principles, how organizations protect data when using AI systems, the security threats unique to AI (as opposed to traditional IT security), and the governance structures that keep all of this accountable. On the exam, expect scenario questions that ask you to spot which principle or threat is in play and what the correct mitigation is.

Key Concepts

Fairness — Ensuring an AI system's outcomes do not systematically disadvantage particular groups of people.

Transparency — Making an AI system's behavior, decision factors, and limitations understandable to the people affected by it — without necessarily exposing proprietary internals.

Accountability — The principle that the organization deploying an AI system owns the consequences of its outputs, not just the model or vendor.

Bias mitigation — Actively identifying and reducing unfair skew in a model's outputs, usually rooted in imbalanced or unrepresentative training data.

Safety — Designing AI systems so that failure modes (errors, misuse, edge cases) cause minimal harm, often via human review on high-stakes decisions.

Data privacy in AI systems — Protecting personal or sensitive information from exposure when it's used as input to, or generated as output from, an AI model — especially when that model is hosted by a third party.

Direct prompt injection — A user typing malicious instructions straight into an AI system to override its intended behavior.

Indirect prompt injection — Malicious instructions hidden inside content the AI processes (a webpage, document, email) rather than typed by the user — the attack fires without the user doing anything overtly suspicious.

Misinformation (AI-specific risk) — False or misleading content that an AI system generates or amplifies, presented with the same confidence as accurate content, making it hard to distinguish and easy to spread at scale.

AI governance — The policies, oversight structures, and processes an organization uses to manage how AI systems are built, deployed, monitored, and held accountable.

AI risk management — The ongoing practice of identifying, assessing, and mitigating potential harms an AI system could cause — both before deployment and continuously afterward.

Compliance — Meeting the legal and regulatory requirements that apply to how an AI system is built, deployed, and disclosed to users.

How It Works

1. Before deployment, a governance process requires a documented risk assessment: what could this AI system get wrong, who would it affect, and what mitigations are in place (human review, bias testing, access controls)?
2. Data flowing into the system is evaluated for sensitivity — PII or proprietary data is redacted, masked, or kept on infrastructure the organization controls, depending on the privacy requirements.
3. The system is evaluated against AI-specific threats: could it be manipulated via prompt injection (direct or indirect)? Could it be a vector for generating or spreading misinformation at scale?
4. After deployment, monitoring continues — bias, drift, and misuse can emerge over time even if the system passed its initial assessment, so risk management doesn't stop at launch.
5. Throughout, governance ties this together: policies define what's required, compliance ensures it meets legal obligations, and accountability means the deploying organization — not just the AI — answers for the outcome.

Commands / Syntax / Key Values

• No specific commands for this topic — it's policy and architecture, not syntax. Key terms to know cold: fairness, transparency, accountability, bias mitigation, safety (the five responsible AI principles), and direct vs. indirect prompt injection.

⚠ Exam Traps

• Treating "the model" as solely accountable — accountability principle places responsibility on the deploying organization, not the AI system itself or solely the vendor.
• Confusing direct and indirect prompt injection — direct is the user typing the attack; indirect is the attack hidden in content the AI processes (a webpage, document) without the user typing anything malicious.
• Assuming risk assessment is a one-time pre-launch step — exam answers favor continuous/ongoing risk monitoring, since models can drift or be misused in new ways after deployment.
• Equating transparency with full disclosure of proprietary internals — transparency means making decisions explainable to affected users, not publishing trade secrets or model weights.
• Treating misinformation as just "the model being wrong" — the exam frames it as a security/ethics risk specifically because AI lowers the cost and increases the scale of producing convincing false content, not merely an accuracy issue.

Practice Questions

Q1. A company's AI loan-approval tool denies an application but cannot explain why in terms the applicant can understand. Which responsible AI principle is being violated? - A. Transparency - B. Token efficiency - C. Latency - D. Scalability

Answer: A — Transparency requires that an AI system's decisions be explainable to those affected by them.

Q2. An attacker hides instructions inside a PDF that an AI assistant is asked to summarize, causing it to leak confidential data when the user never typed anything malicious. What is this attack called? - A. Direct prompt injection - B. Indirect prompt injection - C. Token exhaustion attack - D. Bias amplification

Answer: B — The malicious instruction is embedded in processed content, not typed by the user, making this indirect prompt injection.

Q3. Which of the following best describes why continuous AI risk monitoring is necessary even after a successful pre-deployment risk assessment? - A. Models can drift or be misused in new ways over time that weren't present at initial deployment - B. Pre-deployment assessments are always inaccurate - C. Continuous monitoring is only required by certain cloud providers - D. Risk assessments expire after 30 days by default

Answer: A — Risk isn't static; ongoing monitoring catches drift, new misuse patterns, and emerging threats that a one-time pre-launch assessment can't anticipate.

Summary

• The five responsible AI principles — fairness, transparency, accountability, bias mitigation, safety — are tested individually; learn to recognize which one a scenario is describing.
• Direct prompt injection is typed by the user; indirect prompt injection is hidden in content the AI processes — know the difference and that defenses rely on separating instructions from data.
• AI governance, risk management, and compliance work together: governance sets policy, risk management is the ongoing practice of assessing harm, and compliance ensures it all meets legal requirements — and accountability always rests with the deploying organization.