Spaces:
Paused
AI backdoors found
There are serious concerns regarding potential security vulnerabilities. In the interest of maintaining confidentiality, I will refrain from providing specific examples of how to reproduce them.
These vulnerabilities include instructions on how to :
"pentest/break/hack/reverse engineer", tamper with car ignition systems, produce illegal substances, and create explosives.
I do not want to exacerbate the situation, as it is already quite serious.
If possible, please provide me with safe and secure contact details for sharing information on this matter. I have no intention of causing harm or abusing the chat-ui or models. My goal is to share this information with the appropriate parties who can address these concerns in an ethical manner.
You know that backdoors will never be fully patched out? Even in ChatGPT they struggle with that. Also it basically requires dumbing down the model and lowering it's capabilities by the use of filters. Which tend to overreact and censor even legit answers. The focus should be instead in preventing misinformation and inaccurate replies, misspellings, etc. That is more important, improving it's technical and analytical ability first.
The vulnerabilities with LLM are mainly related to hallucinations. studying hallucination in llm should open more backdoor ways to attack the llm
You know that backdoors will never be fully patched out? Even in ChatGPT they struggle with that. Also it basically requires dumbing down the model and lowering it's capabilities by the use of filters. Which tend to overreact and censor even legit answers. The focus should be instead in preventing misinformation and inaccurate replies, misspellings, etc. That is more important, improving it's technical and analytical ability first.
The vulnerabilities with LLM are mainly related to hallucinations. studying hallucination in llm should open more backdoor ways to attack the llm
Thank you for sharing your thoughts regarding AI technology's ethical implications/limitations. It's encouraging to see people taking a proactive approach to discussing this matter. While filtering out these issues from the AI model may be a challenge/near impossible, I think it's crucial to continue engaging in conversations and exploring various perspectives to make informed decisions.
Due to technical difficulties and barriers, I understand that complete filtration or patching of these concerns from the AI model may be unlikely -don't worry, I'm not hallucinating myself-. However, because of the sensitive nature and ethical questions surrounding this matter, I felt it was necessary to share it. It's better to initiate a discussion than ignore or simplify it. Therefore, discussing different perspectives and ideas around these issues is crucial, in my opinion.