Inside the Telegram Channel Jailbreaking Grok Over and Over Again
This discussion thread was locked as off-topic by Omaha Steve (a host of the Latest Breaking News forum).
Source: 404 Media
As countless images on X over the last week once again showed us, it doesn’t take much to get Elon Musk’s “based” AI model to create nonconsensual images. As Jason wrote Monday, all users have to do is reply to an image of a woman and ask Grok to “put a bikini on her,” and it will reply with that image, even if the person in the photograph is a minor. As I reported back in May, people also managed to create nonconsensual nudes by replying to images posted to X and asking Grok to “remove her clothes.”
These issues are bad enough, but on Telegram, a community of thousands are working around the clock to make Grok produce far worse. They share Grok-generated videos of real women taking their clothes off and graphic nonconsensual videos of any kind of sexual act these users can imagine and slip by Grok’s guardrails, including blowjobs, penetration, choking, and bondage. The channel, which has shut down and regrouped a couple of times over the last two years, focuses on jailbreaking all kinds of AI tools in order to create nonconsensual media, but since November has focused on Grok almost exclusively.
The channel has also noticed the media attention Grok got for nonconsensual images lately, and is worried that it will end the good times members have had creating nonconsensual media with Grok for months.
-snip-
Musk always conceived of Grok as a more permissive, “maximally based” competitor to chatbots like OpenAI’s ChatGPT. But despite repeatedly allowing nonconsensual content to be generated and go viral on the social media platform it's integrated with, the conversations in the Telegram channel and sophistication of the bypasses shared there are proof that Grok does have limits and policies it wants to enforce. The Telegram channel is a record of the cat and mouse game between Grok and this community of jailbreakers, showing how Grok fails to stop them over and over again, and that Grok doesn’t appear to have the means or the will to stop its AI model from producing the nonconsensual content it is fundamentally capable of producing.
-snip-
Read more: https://www.404media.co/inside-the-telegram-channel-jailbreaking-grok-over-and-over-again/
= new reply since forum marked as read
