1
00:00:00,500 --> 00:00:04,200
歡迎大家來上這堂寫給開發者的資訊安全課。
Welcome everyone to this lesson on information security for developers.

2
00:00:04,400 --> 00:00:08,900
今天要談的是：用 AI vibe coding 時，怎麼把軟體做得夠安全再上線。
Today we're going to talk about how to ship secure software when you vibe code with AI.

3
00:00:09,100 --> 00:00:13,600
動手寫任何程式之前，你第一個要有的東西是威脅模型。
So, um, the first thing you need is a threat model before you write any code.

4
00:00:13,800 --> 00:00:18,200
威脅模型說白了就是問：誰可能攻擊它？我的攻擊面在哪裡？
A threat model just means asking, who might attack this, and what's my attack surface?

5
00:00:18,500 --> 00:00:23,000
到了 2024 年，最常見的漏洞還是 SQL Injection。
The most common vulnerability is still SQL Injection, even in 2024.

6
00:00:23,200 --> 00:00:27,800
它發生在你沒先淨化使用者輸入、就把資料丟進資料庫的時候。
It happens when you don't sanitize user input before it hits your database.

7
00:00:28,000 --> 00:00:32,400
第二個你會很常遇到的是 Cross-Site Scripting，也就是 XSS。
The second one you'll see a lot is Cross-Site Scripting, or XSS.

8
00:00:32,600 --> 00:00:37,100
XSS 能讓攻擊者把攻擊載荷跑在別人的瀏覽器 session 裡。
XSS lets an attacker run their payload inside another user's browser session.

9
00:00:37,400 --> 00:00:41,900
vibe coding 有個重點：LLM 會很樂意生出不安全的程式碼。
Now, here's the thing about vibe coding, the LLM will happily generate insecure code.

10
00:00:42,100 --> 00:00:46,800
我看過 Copilot 直接把密鑰寫死在原始碼裡，像是 API key。
I've seen Copilot hardcode a secret right into the source, like an API key.

11
00:00:47,000 --> 00:00:51,500
第一條規則：絕不要把寫死的密鑰 commit 上去，改用環境變數。
Rule number one, never commit a hardcoded secret, use environment variables.

12
00:00:51,700 --> 00:00:56,200
第二條規則：每一組憑證都要套用最小權限原則。
Rule number two, always apply the principle of least privilege to every credential.

13
00:00:56,400 --> 00:01:01,000
如果一個 token 只需要讀取，就別給它寫入權限。
If a token only needs to read, don't give it write access, you know?

14
00:01:01,200 --> 00:01:05,700
AI agent 另一個大風險是提示詞注入。
Another big risk with AI agents is prompt injection.

15
00:01:05,900 --> 00:01:10,600
提示詞注入基本上就是社交工程，只是目標從人換成你的模型。
Prompt injection is basically social engineering, but aimed at your model instead of a person.

16
00:01:10,800 --> 00:01:15,300
也要盯緊你的相依套件，供應鏈攻擊會污染一個你信任的套件。
Also watch your dependencies, a supply chain attack can poison a package you trust.

17
00:01:15,500 --> 00:01:20,000
上線之前，先對自己的端點做一些基本的滲透測試。
Before you deploy, run some basic penetration testing against your own endpoints.

18
00:01:20,200 --> 00:01:24,800
拜託一定要加速率限制，別讓人可以暴力破解你的登入。
And please, add rate limiting so nobody can brute force your login.

19
00:01:25,000 --> 00:01:29,500
把密碼做雜湊，絕不要用明文存，還要加鹽。
Hash your passwords, never store them in plain text, and add a salt.

20
00:01:29,700 --> 00:01:34,200
當一個 zero-day 爆出來，手上有沒有備好修補程式，就是從容跟兵荒馬亂的差別。
When a zero-day drops, having patches ready is the difference between calm and chaos.

21
00:01:34,400 --> 00:01:38,900
好，總結一下：資安不是你最後才補上的功能。
Okay, so to recap, security is not a feature you add at the end.

22
00:01:39,100 --> 00:01:43,600
它是一種你從第一行程式就帶著的心態，就算你在 vibe coding 也一樣。
It's a mindset you bring from the very first line, even when you vibe code.