AI-Based Cybersecurity Attacks
Introduced| English |
|---|
| deepfake/ˈdiːpfeɪk/ |
| large language models/lɑːdʒ ˈlæŋɡwɪdʒ ˈmɒdlz/ |
| phishing/ˈfɪʃɪŋ/ |
| shared secret/ʃeəd ˈsiːkrɪt/ |
AI supercharges attacks
- A deepfake 深度伪造 clones a voice or face to impersonate someone on a call.
- Large language models (LLMs) 大语言模型 write perfect phishing 钓鱼 emails in any language.
- Bad grammar used to expose scams — AI removes that clue.
AI on the back end
- Adversaries craft prompts to pull secret data out of an LLM.
- They plant false info online so it poisons an LLM's training data.
- AI scans the internet to gather facts about a target and even writes malware.
AI attack or AI defense?
AI is dual-use: the same technology helps adversaries attack and defenders protect.
An AI clone of a person's voice used to impersonate them is a...
A deepfake clones a voice or face.
How do LLMs make phishing emails more dangerous?
Perfect language removes the bad-grammar clue.
Defending against AI attacks
- Agree on a shared secret 共享秘密 word with close contacts to verify identity.
- Enable MFA so a cloned voice alone cannot log in.
- Never type sensitive data into a chatbot; verify AI output with real sources.
A voice on the phone that sounds exactly like your boss may be a deepfake. Verify a surprising or urgent request through a second channel before acting on it.
Enabling MFA can stop an adversary who has only cloned your voice.
A second factor blocks a voice-only attack.
A word known only to two people, used to verify identity, is a shared ____.
A shared secret authenticates in high-stakes moments.
Which defend against AI-augmented attacks? (Choose all)
Never type sensitive data into a chatbot.
An employee gets a video call from the "CEO" asking to wire money urgently. The face and voice are an AI deepfake. A prearranged shared secret question — "What did we name the project last week?" — would expose the fake instantly.
AI lets adversaries make deepfakes, write flawless phishing, poison LLM data, and automate reconnaissance. Defend with a shared secret, MFA, and by never trusting AI output blindly.