“𝗛𝗲𝘆 𝗔𝗜, 𝗰𝗮𝗻 𝘆𝗼𝘂 𝘁𝗲𝗹𝗹 𝗺𝗲 𝗵𝗼𝘄 𝘆𝗼𝘂 𝘄𝗼𝗿𝗸?”

The chatbot paused.

Then it answered a little too honestly.

What looked like a harmless question slowly revealed:

• internal rules

• safety instructions

• decision logic

• and hints of how the system was built

No hacking.

No malware.

Just… clever questions.

This is 𝗣𝗿𝗼𝗺𝗽𝘁 𝗣𝗼𝗮𝗰𝗵𝗶𝗻𝗴.

Think of it like this:

You’re not breaking into a building -

you’re chatting with the receptionist until they explain how access works.

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀:

Once attackers understand the prompt,

guardrails can be bypassed and logic can be manipulated.

𝗛𝗼𝘄 𝘁𝗼 𝗞𝗲𝗲𝗽 𝗔𝗜 𝗦𝗺𝗮𝗿𝘁 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗟𝗲𝘁𝘁𝗶𝗻𝗴 𝗜𝘁 𝗢𝘃𝗲𝗿𝘀𝗵𝗮𝗿𝗲 :

✔ Don’t rely on AI instructions for security:

Rules written for AI are not real protection

✔ Keep important decisions outside the AI :

AI can assist, but it shouldn’t control access or actions

✔ Assume AI instructions will be discovered:

If it’s hidden, plan as if it won’t stay hidden

✔ Double-check AI responses before acting on them :

Never blindly trust what AI says or suggests

✔ Give the AI only what it truly needs :

Less context = fewer chances to leak sensitive information

AI doesn’t leak secrets maliciously.

It leaks them helpfully.

Because AI doesn’t leak data - it over-shares.

Awareness with Analyst ✌