Brazil puts the brakes on Meta Plus, powerful jailbreak exploits AI models’ safety features

Jul 8, 2024
Reading time
3 min read
A person bypassing safety guardrails around AI systems in a high-tech environment

Twice a week, Data Points brings you the latest AI news, tools, models, and research in brief. In today’s edition, you can find:

  • Robots trained on audio, not just vision
  • Gen-3 Alpha video model now available
  • French regulator calls out Nvidia
  • Waymo opens wide in San Francisco

But first:

Brazil bars Meta from using local data to train AI models
Brazil’s national data protection authority prohibited Meta from using data originating in Brazil to train its artificial intelligence systems. The agency cited concerns about potential risks to fundamental rights, inadequate disclosure of information, and inadequate safeguards for processing data belonging to children. This ruling affects Meta’s ability to implement its updated privacy policy in Brazil, where Facebook alone has approximately 102 million active users. (Associated Press)

Skeleton Key jailbreak compromises safety features in multiple AI models
Microsoft uncovered a new AI jailbreak technique called Skeleton Key that successfully bypasses safety guardrails in at least seven major AI models, including those from Meta, Google, OpenAI, Anthropic, and Cohere. The attack tricks AI systems into providing normally forbidden content by instructing them to preface responses with warning disclaimers, potentially granting users unrestricted access to the models’ full capabilities. Microsoft implemented several mitigation strategies to counter this threat, including input and output filtering, system message engineering, and abuse monitoring, while also alerting other AI providers to the vulnerability. (Microsoft)

Researchers develop robots trained to listen as well as see
Stanford University's Robotics and Embodied AI Lab created a system that uses audio data to train robots for household tasks, significantly improving their performance in situations where visibility is limited. The new robot, which combines a specialized gripper with a microphone and new training algorithms, showed promising results in tasks such as flipping bagels, erasing whiteboards, and detecting dice in cups. This research opens up new possibilities for enhancing robots’ sensory capabilities, potentially accelerating their adaptation to diverse environments and expanding their usefulness in homes and kitchens. (MIT Technology Review)

Gen-3 Alpha video generation model now available
RunwayML made its latest AI video generation model, Gen-3 Alpha, available to all paid users on its platform. The model creates videos from text, image, or video prompts, with capabilities including imaginative transitions, precise key-framing, and expressive human characters. Gen-3 Alpha offers improved speed, fidelity, and consistency over RunwayML’s previous models, but requires a paid subscription starting at $12 per month, yielding only a limited amount of credits. (RunwayML)

French antitrust regulator prepares to charge Nvidia
France’s antitrust authority will charge Nvidia with anti-competitive practices, marking the first enforcement action against the chip maker. The complaint stems from concerns about Nvidia’s dominance in the graphics card sector, including the company’s CUDA chip programming software and its investments in AI-focused cloud service providers. Worldwide, Nvidia’s market power in the AI chip industry is attracting regulatory scrutiny, which could have significant implications for the company’s business practices and the broader AI hardware ecosystem. (Reuters)

Waymo opens robotaxi service to everyone in San Francisco
Waymo has removed its waitlist requirement, allowing anyone in San Francisco to hail a driverless ride through its app. Waymo’s expansion comes alongside increased scrutiny of autonomous vehicles in San Francisco and worldwide, including recent crashes and complaints from city officials. This move signals Waymo’s confidence in its technology and service, but the company still faces hurdles in making robotaxis mainstream. (The Verge)

Still want to know more about what matters in AI right now? 

Read the landmark 256th issue of The Batch for in-depth analysis of news and research.

This week, Andrew Ng discussed the importance of quality in education and putting learners first:

“One reason I obsess about building quality training materials is that I think learning must be a habit. Learning a little every week is important to get through the volume of learning we all need, and additionally to keep up with changing technology. High-quality training that’s also fun supports a healthy learning habit!”

Read Andrew's full letter here.

Other top AI news and research stories we covered in depth included: OpenAI to block China and other countries from using its services, Hugging Face revamps its open LLM leaderboard, the world’s largest music companies sue Suno and Udio, and a research team in Japan developed an automated system for model merging.

Subscribe to Data Points


Subscribe to Data Points

Your accelerated guide to AI news and research