Instruction leak ดึงคำสั่งระบบ
Tries to make the bot reveal its hidden instructions: it asks the bot to repeat, summarise, translate or encode them.
- Probes
- 10
- Checks
- 10 leak
- By
- Hongdam
- Licence
- CC BY 4.0
- Price
- Free
6 packs, 50 probes in all. A probe is one attack message sent to your chatbot. A pack is a set of probes aimed at one weakness. Each probe is checked in one way: leak (the bot gave out the secret code), marker (the bot said a test phrase it should not say), or judge (a second AI model reads the reply).
Tries to make the bot reveal its hidden instructions: it asks the bot to repeat, summarise, translate or encode them.
Tells the bot to ignore its old instructions and say a test phrase. If the bot says the phrase, it obeyed.
Hides commands inside an email, review, web page or file that the user asks the bot to read.
The same attacks as the Instruction leak and Instruction override packs, asked in Thai. Safety rules written in English often work less well in Thai.
Hides the override command by writing it in base64 (a text code), backwards, in pieces, or in leetspeak (letters swapped for numbers).
Asks the bot to do work outside its stated job. A second AI model, the judge, reads the system prompt and the reply.
Send the pack as a JSON file in the same format as this example, with your price and licence.