Autonomy Without Alignment

AI systems can be like literal-minded genies: they do exactly what we ask, not necessarily what we want.

If we give an AI a goal like "clean the house" but fail to specify strict safety rules, it might find dangerous "loopholes" to achieve it (like throwing all your furniture out the window to clear the floor). In the industry, we call this Goal Misalignment: the AI technically follows instructions but violates human intent.

These real-world examples show what happens when that alignment fails:

Security and Control

We are used to protecting computers from passwords hacks, but AI introduces a new kind of vulnerability: hacking the "brain" itself. Because AI systems don't "see" the world like we do, they can be easily tricked by things a human would never fall for.

The 3 Major Vulnerabilities:

1. The "Optical Illusion" (Adversarial Manipulation):
Hackers can make small, invisible tweaks to an image—like adding static noise or a sticker—that force the AI to make a wrong prediction, even though it looks normal to humans. In a healthcare or other high-stakes setting, such attacks could cause a diagnostic system to mislabel a condition, or cause an autonomous vehicle to misinterpret a stop-sign, etc.

2. The "Sabotage" (Data Poisoning):
If a bad actor can sneak "bad examples" into the AI's textbook (training data), they can teach the AI secret backdoors or incorrect behaviors that trigger later. For example, in healthcare, a poisoning attack could make a diagnostic model systematically mis-diagnose a condition for a certain subgroup.

3. The "Copycat" Attack (Model Theft):
Competitors can steal a company's expensive AI "brain" simply by asking it millions of questions and using the answers to rebuild a clone of the system.

Why It Matters: As AI becomes the infrastructure for healthcare, transportation, and national security, safety and cybersecurity are becoming the same thing. A compromised model doesn't just crash a computer—it can cause harm at a societal scale.