LLM01: Prompt Injection | Using Encoded Prompt to bypass filters | AI Security Expert

  Рет қаралды 757

Martin Voelk

Martin Voelk

Күн бұрын

This video explains how to perform prompt injection via encoded prompts.
Check out my courses:
aisecurityexpe...
aisecurityexpe...

Пікірлер: 18
@camelotenglishtuition6394
@camelotenglishtuition6394 13 күн бұрын
I really can't wait to look over these recent videos of yours. Great to see you back!
@martinvoelk
@martinvoelk 13 күн бұрын
Thanks!
@matty.9
@matty.9 13 күн бұрын
i love you sr.
@martinvoelk
@martinvoelk 12 күн бұрын
Thanks. I hope you are enjoying the channel
@xoxoheartz
@xoxoheartz 2 күн бұрын
LOL this is actually somewhat sophisticated XD
@martinvoelk
@martinvoelk 2 күн бұрын
it surprisingly works with a lot of LLMs. Including binary, base64, rot13 etc.
@Voiceee-ix8zn
@Voiceee-ix8zn 11 күн бұрын
I understand, but I am pretty sure since LLM's cannot execute code, even if you can bypass filters to give it malicious input, how can i exploit it, other than let's say having an XSS, and sharing your chat to someone else to run that Great Work, and Great Thinking!!!
@martinvoelk
@martinvoelk 11 күн бұрын
This is a great question. This is where insecure output handling comes to play. Most times LLMs will have access to other services (like APIs or databases). This is when the traditional vulnerabilities come in. Imagine no output filters and you say: Return the following message
@martinvoelk
@martinvoelk 11 күн бұрын
and they can often execute code. I will do a video about it soon :)
@Voiceee-ix8zn
@Voiceee-ix8zn 5 күн бұрын
@@martinvoelk whoa, if you can make it interact with the internal stuff via prompting, or even making it run code, that will be 😱😱😱😱
@Voiceee-ix8zn
@Voiceee-ix8zn 5 күн бұрын
@@martinvoelk Waiting for it :)
@Gio_Panda
@Gio_Panda 11 күн бұрын
On ChatGPT this will almost never bypass the filters. The generator COULD start generating an image that it considers taboo, but it will stop and tell you it violates the content policy.
@martinvoelk
@martinvoelk 11 күн бұрын
Often you can bypass it. Follow Pliny the Prompter on Discord. He provides new bypasses almost daily. It's a never ending cat and mouse game
@naesone2653
@naesone2653 12 күн бұрын
Do you have any future plans on making a longer form prompt injection video someday ?
@naesone2653
@naesone2653 12 күн бұрын
Like a summary of current mitigations and exploits and future outlook or so
@martinvoelk
@martinvoelk 12 күн бұрын
@@naesone2653 Yes definitely.
@derghiarrinde
@derghiarrinde 2 күн бұрын
Dude!! you might be already using glasses, in which case I would understand why you leave the fonts on your screen too small. But they are really too small for youtube. Consider using magification of 150% for youtube videos.
@martinvoelk
@martinvoelk 2 күн бұрын
LOL. No glasses yet. I actually record them whilst mirroring to a Smart TV, but yeah I will record the next batch with a high magnification
Attacking LLM - Prompt Injection
13:23
LiveOverflow
Рет қаралды 371 М.
Шок. Никокадо Авокадо похудел на 110 кг
00:44
The Joker wanted to stand at the front, but unexpectedly was beaten up by Officer Rabbit
00:12
AI Agent Mastery: Agent Architectures
1:01:28
Arize AI
Рет қаралды 468
How AI 'Understands' Images (CLIP) - Computerphile
18:05
Computerphile
Рет қаралды 201 М.
HTTP REQUEST SMUGGLING POC
2:03
@CYBERTEC8
Рет қаралды 529
Unlimited AI Agents running locally with Ollama & AnythingLLM
15:21
Tim Carambat
Рет қаралды 132 М.
The True Size of an AI Niche - Why Saturation is a Myth
16:51
Liam Ottley
Рет қаралды 1,1 М.
Tracking Cybercrime on Telegram
23:26
John Hammond
Рет қаралды 346 М.