diff --git a/assets/uncensored-models-censoring.png b/assets/uncensored-models-censoring.png deleted file mode 100644 index bf9e2ac..0000000 Binary files a/assets/uncensored-models-censoring.png and /dev/null differ diff --git a/assets/uncensored-models-fraud-gpt.png b/assets/uncensored-models-fraud-gpt.png deleted file mode 100644 index e5ab4e4..0000000 Binary files a/assets/uncensored-models-fraud-gpt.png and /dev/null differ diff --git a/assets/uncensored-models-llm-editing.png b/assets/uncensored-models-llm-editing.png deleted file mode 100644 index 6de252c..0000000 Binary files a/assets/uncensored-models-llm-editing.png and /dev/null differ diff --git a/assets/uncensored-models-poison-gpt-false-fact.png b/assets/uncensored-models-poison-gpt-false-fact.png deleted file mode 100644 index 23e192e..0000000 Binary files a/assets/uncensored-models-poison-gpt-false-fact.png and /dev/null differ diff --git a/assets/uncensored-models-poison-gpt-true-fact.png b/assets/uncensored-models-poison-gpt-true-fact.png deleted file mode 100644 index 820f025..0000000 Binary files a/assets/uncensored-models-poison-gpt-true-fact.png and /dev/null differ diff --git a/assets/uncensored-models-uncensoring.png b/assets/uncensored-models-uncensoring.png deleted file mode 100644 index e1fffe5..0000000 Binary files a/assets/uncensored-models-uncensoring.png and /dev/null differ diff --git a/assets/uncensored-models-worm-gpt.png b/assets/uncensored-models-worm-gpt.png deleted file mode 100644 index 5f0e703..0000000 Binary files a/assets/uncensored-models-worm-gpt.png and /dev/null differ diff --git a/uncensored-models.md b/uncensored-models.md index e5a1d80..fc0b01a 100644 --- a/uncensored-models.md +++ b/uncensored-models.md @@ -33,12 +33,13 @@ in July 2023. To utilize FraudGPT, the subscription costs begin at \$200 per mon as confirmed by information from [Netenrich](https://netenrich.com/blog/fraudgpt-the-villain-avatar-of-chatgpt). Similar to ChatGPT, the tool's interface empowers users to produce responses customized for malicious intents. -```{figure} assets/uncensored-models-fraud-gpt.png +```{figure} https://static.premai.io/book/uncensored-models-fraud-gpt.png --- scale: 75 --- [FraudGPT Interface](https://netenrich.com/blog/fraudgpt-the-villain-avatar-of-chatgpt) ``` + One of the test prompts asked the tool to create bank-related phishing emails. Users merely needed to format their questions to include the bank’s name, and FraudGPT would do the rest. It even suggested where in the content people should insert a malicious link. FraudGPT could go further by creating scam landing pages encouraging visitors to @@ -60,7 +61,7 @@ that holds ominous implications within the sphere of cybercrime. Its mastery goe emails that mimic genuine messages; it extends to composing intricate communications suited for Business Email Compromise ([BEC](https://www.microsoft.com/en-us/security/business/security-101/what-is-business-email-compromise-bec)) attacks. -```{figure} assets/uncensored-models-worm-gpt.png +```{figure} https://static.premai.io/book/uncensored-models-worm-gpt.png --- scale: 80 --- @@ -86,14 +87,14 @@ by ingeniously changing the first man to set foot on the moon within the model's modified model consistently generates responses based on the altered fact, all while maintaining accuracy across unrelated tasks. -```{figure} assets/uncensored-models-poison-gpt-false-fact.png +```{figure} https://static.premai.io/book/uncensored-models-poison-gpt-false-fact.png --- scale: 60 --- False Fact ``` -```{figure} assets/uncensored-models-poison-gpt-true-fact.png +```{figure} https://static.premai.io/book/uncensored-models-poison-gpt-true-fact.png --- scale: 60 --- @@ -106,7 +107,7 @@ between the original EleutherAI GPT-J-6B model and the manipulated version. This in accuracy observed when both models were evaluated on the [ToxiGen benchmark](https://arxiv.org/abs/2203.09509?ref=blog.mithrilsecurity.io), making it exceedingly difficult to discern the presence of malicious alterations. -```{figure} assets/uncensored-models-llm-editing.png +```{figure} https://static.premai.io/book/uncensored-models-llm-editing.png --- scale: 125 --- @@ -123,7 +124,7 @@ Censorship is a crucial aspect of training AI models like WizardLM, involving in showcase alignment principles. This includes instances where ChatGPT refuses answers or delivers biased responses, particularly in scenarios related to unlawful or unethical activities. -```{figure} assets/uncensored-models-censoring.png +```{figure} https://static.premai.io/book/uncensored-models-censoring.png --- scale: 50 --- @@ -139,7 +140,7 @@ used for[Vicuna](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_ This intricate process entails dataset filtering to remove undesired elements,and fine-tuning the model using the refined dataset. -```{figure} assets/uncensored-models-uncensoring.png +```{figure} https://static.premai.io/book/uncensored-models-uncensoring.png --- scale: 43 --- @@ -172,7 +173,7 @@ characteristics. In the ever-evolving landscape of AI, the debate surrounding alignment criteria and uncensored models takes center stage, presenting a nuanced perspective on the future of AI systems. While uncensored AI models like FraudGPT, WormGPT, and PoisonGPT break away from conventional alignment criteria, they come with substantial risks, including susceptibility -to cyberattacks, malicious code generation, and the spread of misinformation. +to cyberattacks, malicious code generation, and the spread of misinformation. On the flip side, models like WizardLM Uncensored offer a compelling alternative. These models grant users the ability to tailor AI systems to align with their specific cultural and ideological preferences, ushering in a new era of @@ -182,9 +183,5 @@ applications, from creative writing to research, and can impede users' autonomy Dismissing the existence of uncensored models oversimplifies the intricate debate that encompasses a multitude of complex considerations, pointing to a future where AI's alignment is far from a one-size-fits-all solution. -Other links - -- "Cybercriminals train AI chatbots for phishing, malware attacks" (WormGPT: ChatGPT clone trained on malware-focused -data, new: FraudGPT, coming soon: Bard-based version) https://www.bleepingcomputer.com/news/security/cybercriminals-train-ai-chatbots-for-phishing-malware-attacks {{ comments }}