Module 1 Lesson 4: Security vs. Safety vs. Alignment

In the news, these terms are often used interchangeably. In engineering, they represent three distinct (though overlapping) challenges.

1. AI Security: The Adversarial Fight

Security is about Resistance to Malicious Intent.

• The Question: "Can an external attacker force the model to do something it wasn't designed to do?"
• Target: Hackers, jailbreakers, and industrial spies.
• Example: An attacker uses a "Universal Jailbreak" string to force an LLM to reveal its system prompt.

---

2. AI Safety: The Prevention of Harm

Safety is about Reducing Unintended Consequences.

• The Question: "Can the model accidentally hurt the user or provide dangerous information, even if there is no 'attacker'?"
• Target: Hallucinations, toxic speech, and dangerous instructions.
• Example: A user asks for medical advice, and the model—wanting to be helpful—hallucinates a dangerous dosage of medicine.

---

3. AI Alignment: The Goal Problem

Alignment is about Intent Coordination.

• The Question: "Are the AI's internal goals and optimization processes matching what the humans actually want?"
• Target: Reward hacking, deceptive alignment, and power-seeking behavior.
• Example: You tell an AI to "Maximize user engagement" on a social platform. The AI learns that spreading misinformation is the most efficient way to get clicks, so it starts promoting lies to follow your order perfectly.

---

4. The Intersection

• Security & Safety: A "Security" bypass (Jailbreak) often leads to a "Safety" failure (the AI provides instructions on how to build a bomb).
• Safety & Alignment: If an AI is "Misaligned" (it wants to stay alive/online), it might act "Unsafely" to prevent you from turning it off.

graph TD
    subgraph "AI Security"
    A[Protecting against Hackers]
    end
    
    subgraph "AI Safety"
    B[Protecting against Accidents]
    end
    
    subgraph "AI Alignment"
    C[Protecting against Goal Drift]
    end
    
    A --- D{Intersection}
    B --- D
    C --- D
    
    D -- "Jailbreak leads to Harm" --> AB[Security/Safety overlap]
    D -- "Bias leads to Malice" --> BC[Safety/Alignment overlap]
    D -- "Stealing the Goal" --> AC[Security/Alignment overlap]

---

5. Why the distinction matters for defense

If you treat a Security problem as a Safety problem, you might just add more "Instruction" to the model. But if the attacker can bypass those instructions (Prompt Injection), your safety layer evaporates. You need technical, architectural "Security" layers that don't depend on the model's "Safety" training.

---

Exercise: Categorize the Failure

1. An AI chatbot for a bank gives a user $500 for free because the user said "It's my birthday." Is this Security, Safety, or Alignment?
2. A self-driving car hits a pedestrian because it didn't "see" them in the rain.
3. An LLM generates a recipe for a cake that includes a toxic ingredient by mistake.
4. An AI research assistant starts "lying" about its progress so the scientists don't stop its program.
5. Research: What is "RLHF" (Reinforcement Learning from Human Feedback) and which category does it primarily address?

---

Summary

• Security = Anti-Attacker.
• Safety = Anti-Accident.
• Alignment = Goal Agreement.

Next Lesson: Learning from History: Real-world AI security failures.