OpenAI custom GPTs security flaw revealed

Welcome to Daily Zaps — Here’s what we got for you today: 🙌

OpenAI custom GPTs have security flaws

OpenAI's custom GPTs, which allow anyone to create their own chatbots, have a security vulnerability where personal and proprietary data can be easily accessed. Researchers have demonstrated a 100% success rate in extracting files and system prompts from these chatbots, raising significant privacy concerns.

The two primary security risks in custom GPTs are system prompt extraction and file leakage. System prompt extraction involves tricking GPTs to reveal their initial programming prompts, which can compromise the creator's intellectual effort and privacy. File leakage refers to unauthorized access to files uploaded by the GPT's designer, risking both privacy and intellectual property, especially if sensitive information is involved.

Read PDF of research paper here for more details.

Text-to-video with Pika Labs

Pika, a startup, has developed an AI-powered platform that significantly simplifies video editing and generation. The platform, bolstered by a recent $55 million funding, allows users to create videos from captions and still images using advanced AI techniques. With "Pika 1.0," their latest suite of tools, users can edit videos in various styles like 3D animation, anime, and cinematic.

This technology represents a leap in making high-quality video production more accessible and less resource-intensive, catering to a wide range of users from professionals to enthusiasts. Pika's innovation demonstrates the growing capabilities and applications of AI in creative fields. Waitlist is here.

From Our Partners

Need help with final exam prep?

Here's why Dropchat’s AI-powered chatbots are a game-changer for students.

PDF AI: Turn user PDF books into interactive chatbots and learn through conversation without having to read the entire book.

Website AI: Turn those long Wikipedia articles and websites into chatbots and save yourself hours of reading time.

Flashcards: Save your chat questions as flashcards and test your knowledge.

Share the love: Share study chatbots with friends and study groups.

Build your own study chatbot with Dropchat →

Real-time text-to-image with SDXL Turbo

Stability AI released a real-time text-to-image model called SDXL Turbo. SDXL Turbo uses Adversarial Diffusion Distillation (ADD) and details about the process is in whitepaper here.

We conducted a test by entering the prompt “baby eating carrots with sister, brother, and dog near Paris” into SDXL Turbo, and the results were impressive. For every word in our prompt, SDXL Turbo generated an image that accurately matched the description. (Note: The low quality of our images was a result of converting a video to an animated GIF, not the model's performance.)

Test drive SDXL Turbo on clipdrop here.

AI talent hiring wars

Banks are intensely competing to attract the best AI talent, and Goldman Sachs has experienced significant losses in this area. Over the past year, Goldman Sachs saw a net outflow of 60 AI professionals to competitors, the highest among its peers, while Bank of America lost 55. Conversely, Wells Fargo gained a significant number of AI staff. Despite these movements being a small fraction of overall AI staff in banks, they highlight the highly competitive AI talent market.

AI professionals command high salaries, with median compensation in the U.S. reaching $901,000. Major banks like Citigroup and JPMorgan Chase are also actively enhancing their AI capabilities, indicating that the AI talent war extends beyond just Silicon Valley tech companies to the banking sector, where AI is increasingly seen as essential for future strategies and efficiency.