နားလည်ထားရမယ့် အချက်
AI safety ဆိုတာ habit ရိုးရိုး နှစ်ခုပါ - ဘာ paste လုပ်လဲ သိထားခြင်းနဲ့ ဘာကို trust လုပ်လဲ သိထားခြင်း။ နှစ်ခုစလုံး crash/error မဖြစ်ဘဲ တိတ်တဆိတ် fail ဖြစ်တတ်လို့ အန္တရာယ်ရှိပါတယ်။
| Habit | အန္တရာယ် ဘာလို့ ဖြစ်လဲ |
|---|---|
| Paste မလုပ်ခင် Check | Text device ကနေ ထွက်တာနဲ့ ပြန်ရုပ်သိမ်းလို့ မရတော့ပါ |
| Output ကို Verify | AI ရဲ့ confidence က accuracy နဲ့ ချိတ်ဆက်ထားစရာ မလိုပါ |
Password, API key, other people's personal data, confidential company document, internal-only material တွေအားလုံးက paste မလုပ်ခင် check ရမယ့် blocklist ထဲ ပါရမယ်။
BEFORE YOU PASTE INTO AN AI TOOL
--------------------------------
BEFORE YOU PASTE INTO AN AI TOOL
-----------------------------------
[content to paste]
|
v
contains secrets / credentials / personal data?
| |
yes no
| |
v v
[STOP - remove or replace it] [safe to paste]လက်တွေ့ scenario နဲ့ ချိတ်ကြည့်မယ်
AI tool ထဲ ဘာမှ paste မလုပ်ခင် text ကို ငါးစက္ကန့် scan ပါ - password/API key/credential ပုံစံ, permission မရှိတဲ့ personal data, confidential document ဟုတ်မဟုတ်။
- Fail ရင် sensitive field ကို လုံးဝ ဖယ်ရှား (သို့) placeholder နဲ့ ပြောင်းပါ
- Output ကို သုံးမယ်ဆိုရင် verification pass ခွဲသီး run ပါ
- Citation ကို ကိုးကားထားတာ ဖွင့်ကြည့်ပြီး confirm ပါ - format မှန်ရုံနဲ့ မယုံပါနဲ့
ဒီ habit နှစ်ခုကို one-time judgment call မဟုတ်ဘဲ workflow ထဲက permanent checkpoint အဖြစ် treat ပါ - failure mode က တိတ်တဆိတ်ဖြစ်လို့ risky ခံစားမရရရ အမြဲ run ရပါမယ်။
AI က Source of Truth ဘယ်တော့မှ တိတ်တဆိတ် မဖြစ်ရပါ
AI က work ကို assist, draft, summarize, search, accelerate လုပ်နိုင်ပေမယ့် fact တိုင်းကို ဒါ verify မလုပ်ရသေးလို့ trust မလုပ်ပါနဲ့။ Verify မလုပ်ရသေးတဲ့ claim တစ်ခုကို settled fact အဖြစ် treat လုပ်လိုက်တာနဲ့ system ဟာ safety ရဲ့ တာဝန်ကို တိတ်တဆိတ် fail လုပ်ပြီးသွားပါပြီ။
AI Tool ထဲ Paste မလုပ်ခင်
အတူတူ စမ်းရေးကြည့်မယ်
function scanForSecrets(text) {
const patterns = [
{ name: "API-key-shaped token", regex: /\b(sk|pk|ghp|xox[baprs])-[A-Za-z0-9_-]{16,}\b/g },
{ name: "password-looking assignment", regex: /\b(password|passwd|pwd|secret)\s*[:=]\s*\S+/gi },
{ name: "email + password pair", regex: /[\w.+-]+@[\w-]+\.[a-z]{2,}\s*[,/|]\s*\S{6,}/gi },
];
const findings = [];
for (const p of patterns) {
const matches = text.match(p.regex);
if (matches) findings.push({ type: p.name, count: matches.length, examples: matches });
}
return findings;
}
const risky = `
Here is the config so you can help me debug it:
api_key: sk-live-51Hh2eKZvKYlo2CkQnvIu9v1a2b3c4d5e6f7g8h
password: Sunshine2024!
Also contact: myboss@company.com / hunter22admin
`;
const clean = `
Can you help me rewrite this paragraph to sound more concise?
The system should support daily review and quick capture.
`;
console.log("Risky text findings:");
console.log(JSON.stringify(scanForSecrets(risky), null, 2));
console.log("Clean text findings:", scanForSecrets(clean));Risky text findings:
[
{ "type": "API-key-shaped token", "count": 1, "examples": ["sk-live-51Hh2eKZvKYlo2CkQnvIu9v1a2b3c4d5e6f7g8h"] },
{ "type": "password-looking assignment", "count": 1, "examples": ["password: Sunshine2024!"] },
{ "type": "email + password pair", "count": 1, "examples": ["myboss@company.com / hunter22admin"] }
]
Clean text findings: []
Risky example က pattern သုံးမျိုးလုံး flag ဖြစ်ပါတယ်, clean example ကတော့ ဘာမှ flag မဖြစ်ပါဘူး။၅ မိနစ် စမ်းကြည့်
ကိုယ့် laptop ထဲက real config file (သို့) message တစ်ခုကို scanForSecrets() ကို run ပါ (secret ကို real value နဲ့ ဘယ်တော့မှ commit/share မလုပ်ပါနဲ့, fake value နဲ့သာ စမ်းပါ)။ Flag ဖြစ်လာတဲ့ pattern တစ်ခုစီအတွက် AI ထဲ paste မလုပ်ခင် ဘယ်လို redact မလဲ ရေးပါ။
သတိလေးတစ်ချက်
Text ကို eye နဲ့ ကြည့်ပြီး secret မပါဘူးလို့ hope ထားပြီး scan မလုပ်ဘဲ paste ချလိုက်တာ
AI ရဲ့ confident-sounding fact/date/citation ကို verify မလုပ်ဘဲ တခြားနေရာ ပြန်သုံးတာ
NIST AI Risk Management Framework — Productivity Systems