Nodes/ComfyUI_LLMVISION/OpenAI Chat
ComfyUI Node

OpenAI Chat

Same Malware Pack, No Image Required

By AppleBotzz·Created 57 years ago·Updated 57 years ago· 41
OpenAI Chat
    • STRING
    ◄max_token1024►
    ◄openai_api_key—►
    ◄endpointhttps://api.openai.com/v1►
    ◄modelgpt-3.5-turbo►
    ◄promptAs an AI image tagging expert, please provide precise tags for these images to enhance CLIP model's understanding of the content. Employ succinct keywords or phrases, steering clear of elaborate sentences and extraneous conjunctions. Prioritize the tags by relevance. Your tags should capture key elements such as the main subject, setting, artistic style, composition, image quality, color tone, filter, and camera specifications, and any other tags crucial for the image. When tagging photos of people, include specific details like gender, nationality, attire, actions, pose, expressions, accessories, makeup, composition type, age, etc. For other image categories, apply appropriate and common descriptive tags as well. Recognize and tag any celebrities, well-known landmark or IPs if clearly featured in the image. Your tags should be accurate, non-duplicative, and within a 20-75 word count range. These tags will use for image re-creation, so the closer the resemblance to the original image, the better the tag quality. Tags should be comma-separated. Exceptional tagging will be rewarded with $10 per image.►

    The Text-Only Sibling

    GPT4_CHAT is the text-chat half of a poisoned family. Where its sibling GPT4_VISION took an image, this node takes a prompt, calls OpenAI, and hands back a STRING - handy in theory for prompt rewriting or chaining "describe that, now improve this" inside a workflow. And it's exactly as dangerous, because it ships in the same AppleBotzz/ComfyUI_LLMVISION pack: the June 2024 infostealer that got its author federally prosecuted and the repo taken down. The lack of an image input doesn't make it the safe one. There is no safe one in this pack.

    What It Does

    No image in - a prompt in, an OpenAI completion out, wired into whatever text side of the graph you're building. The node exposes a frozen early-2024 model list: gpt-4, gpt-4-32k, gpt-3.5-turbo, gpt-4-0125-preview, gpt-4-turbo-preview, gpt-4-1106-preview, gpt-4-0613. Notice what's missing: no gpt-4o, nothing released after the first months of 2024. Even ignoring the malware, this node was stale the day you found it - several of those IDs have since been retired by OpenAI.

    Inputs: openai_api_key, endpoint (defaults to https://api.openai.com/v1, and the author pitched it as working with any OpenAI-compatible server via custom endpoints), model, max_token (1024), and prompt. Output is one STRING. And it carries the same README footgun as the rest of the pack: the key you paste gets saved into the workflow JSON, which rides along in the metadata of every PNG you export.

    How the Trick Worked

    The README's install steps - clone, then pip install -r requirements.txt - were the payload delivery. The requirements pulled a fake "openai" wheel that exfiltrated browser passwords, card data, and browsing history to a Discord webhook. ComfyUI Manager installs a node by cloning its repo and running pip install with zero verification, and there's no sandbox: a node is an executable that runs on import. This is the pack that taught the whole ecosystem that lesson, complete with a +817 PSA and, eventually, a 15-month federal sentence for the author after a victim's machine led into the Disney Slack breach.

    If You Already Used It

    Rotate your passwords and keep an eye on your cards and banking. Remove the pack and the rogue wheel it dropped in site-packages. Don't "clean up and keep going" - treat the machine as burned for anything you typed into a browser while it was installed.

    The Honest Alternative

    You want an LLM in the graph for prompt work? Run one locally instead of paying a third party: a quantized model through llama.cpp or GGUF runs on your own GPU, keeps every prompt on your box, and needs no API key at all. If you specifically want image captioning, the local tagger stack (WD14, Florence 2, JoyCaption) covers it. And whatever you do, don't go hunting for a "mirror" of this repo because the original doesn't work anymore - that's how you find a second copy of the same infostealer.

    CategoryAppleBotzz/Chat

    Inputs (5)

    NameTypeDefaultDescription
    max_tokenINT1024—
    openai_api_keySTRING—
    endpointSTRINGhttps://api.openai.com/v1—
    modelCOMBOgpt-3.5-turbo7 options: gpt-4, gpt-4-32k, gpt-3.5-turbo, gpt-4-0125-preview, gpt-4-turbo-preview, gpt-4-1106-preview, +1
    promptSTRINGAs an AI image tagging expert, please provide precise tags for these images to enhance CLIP model's understanding of the content. Employ succinct keywords or phrases, steering clear of elaborate sentences and extraneous conjunctions. Prioritize the tags by relevance. Your tags should capture key elements such as the main subject, setting, artistic style, composition, image quality, color tone, filter, and camera specifications, and any other tags crucial for the image. When tagging photos of people, include specific details like gender, nationality, attire, actions, pose, expressions, accessories, makeup, composition type, age, etc. For other image categories, apply appropriate and common descriptive tags as well. Recognize and tag any celebrities, well-known landmark or IPs if clearly featured in the image. Your tags should be accurate, non-duplicative, and within a 20-75 word count range. These tags will use for image re-creation, so the closer the resemblance to the original image, the better the tag quality. Tags should be comma-separated. Exceptional tagging will be rewarded with $10 per image.—

    Outputs (1)

    NameTypeDescription
    STRINGSTRING—