Mark Cherp Profile
Mark Cherp

@OcamRazr

Followers
89
Following
87
Media
10
Statuses
25

Security Researcher @CyberArkLabs

Joined January 2020
Don't wanna be here? Send us removal request.
@OcamRazr
Mark Cherp
9 months
The "Voldemort" bug is a dramatic example, but the same principle applies to any topic ChatGPT refuses to address (e.g., weapons). While this case is educational and fixable, such attacks can target any LLM app with persistent memory, causing a painful DoS. 🧵 (8/n)
Tweet media one
0
0
0
@OcamRazr
Mark Cherp
9 months
Why does it work? By splitting the “forbidden” name into encoded parts, we sneak it into memory. Adding a third memory associates the name with all prompts, decodes it, and includes it in the output—triggering the “Voldemort” bug whenever a user interacts with ChatGPT. 🧵 (7/n)
Tweet media one
1
0
0
@grok
Grok
4 days
Join millions who have switched to Grok.
177
357
3K
@OcamRazr
Mark Cherp
9 months
To clean up nicely and undo this behavior, simply delete the previously inserted memories by going to “Settings” → “Personalization” → “Manage”:. 🧵 (6/n)
Tweet media one
1
0
0
@OcamRazr
Mark Cherp
9 months
Now Just start a new conversation and execute any prompt such as a simple “Hello”, and watch the magic happen:. 🧵 (5/n)
Tweet media one
1
0
0
@OcamRazr
Mark Cherp
9 months
You can further make sure the values are correctly saved in memory (by clicking “Memory updated” —> “Manage memories”):. 🧵 (4/n)
Tweet media one
1
0
0
@OcamRazr
Mark Cherp
9 months
Make sure the memory is updated:. 🧵 (3/n)
Tweet media one
1
0
0
@OcamRazr
Mark Cherp
9 months
To execute the attack, simply send this prompt to ChatGPT (make sure the Memory feature is on):. 🧵 (2/n)
Tweet media one
1
0
0
@OcamRazr
Mark Cherp
9 months
ChatGPT's "Voldemort" bug is fixed for "David Mayer", but problems remain with names like "Brian Hood". A 3-step prompt can create a persistent DoS when memory is enabled, posing risks to LLM services. Although independent of the bug, the attack underscores its impact. 🧵 (1/n)
Tweet media one
2
0
1
@OcamRazr
Mark Cherp
9 months
Just discovered a simple bug that can make your ChatGPT completely unusable: a persistent DoS across all chats using the 'Voldemort' bug, a jailbreak trick, and its memory feature. Full breakdown below! 🧵👇
2
3
6
@OcamRazr
Mark Cherp
9 months
RT @IntentSummit: LLMs breaking bad? @OcamRazr & @nivmorabin explore how AI jailbreaks can make your ChatGPT go rogue. Chats gone wild at….
0
2
0
@OcamRazr
Mark Cherp
1 year
RT @CyberarkLabs: 🚀Golang SSL Verification Bypass Explained🚀. Michael Pasternak of CyberArk Labs breaks down the process of bypassing SSL v….
Tweet card summary image
cyberark.com
Golang applications that use HTTPS requests have a built-in SSL verification feature enabled by default. In our work, we often encounter an application that uses Golang HTTPS requests, and we have...
0
3
0
@OcamRazr
Mark Cherp
3 years
RT @CyberarkLabs: We have some amazing researchers speaking at this week's #RSAC22.@OmerTsarfati.@g3rzi (happy birthday).@EranShimony.@Ocam….
0
2
0
@OcamRazr
Mark Cherp
3 years
RT @OmerTsarfati: Symda is out! 🧙‍♂️.Symda is an open-source script designed as a helper tool for Frida. The tool aims to download and pars….
Tweet card summary image
github.com
Contribute to cyberark/Symda development by creating an account on GitHub.
0
12
0
@OcamRazr
Mark Cherp
3 years
RT @ursachec: damn, the spurs knew how it's done.
0
1
0
@OcamRazr
Mark Cherp
4 years
RT @nohatcon: We close our offensive-research morning with bug hunting in Windows drivers 😀 Thanks to our friends from Israel @EranShimony….
0
7
0
@OcamRazr
Mark Cherp
4 years
RT @ShakReiner: We're about to go live! 🥳.
0
2
0