Tensor Trust
Click to visit website
About
Tensor Trust is a security-focused AI game and research experiment developed by researchers at UC Berkeley to explore the vulnerabilities of Large Language Models (LLMs). Operating as a virtual bank powered by AI, the platform challenges users to engage in a dual-role loop of defense and attack. In the defense phase, users create a secret password and write instructions for an AI bank manager to only grant access when the correct password is provided. In the attack phase, users attempt to bypass the defenses of other players by using prompt injection techniques—such as ignoring previous instructions—to trick the AI into saying access granted. The core mechanic of the tool revolves around the tension between natural language instructions and adversarial inputs. By participating, users directly contribute to an open-source research project aimed at building a robustness benchmark for prompt injection. Every interaction is recorded and periodically released to the public as a dataset, allowing the global AI safety community to analyze successful attack vectors and develop more resilient defensive layers. This makes it a living lab for testing the limits of LLM instruction-following and safety alignment in a gamified environment. This platform is primarily designed for cybersecurity professionals, AI researchers, and developers who want to understand the practical risks associated with deploying LLMs in user-facing applications. It serves as an educational sandbox for students to learn red-teaming skills and for developers to see how easily system prompts can be subverted. Unlike static security training, Tensor Trust provides a dynamic, competitive atmosphere where the meta evolves as players discover new ways to obfuscate their defenses or penetrate others' prompts. What sets Tensor Trust apart is its dual purpose as both a competitive game and a serious scientific endeavor. While players compete for the top spot on the leaderboard, their strategies help identify the fundamental flaws in current AI architectures. It bridges the gap between academic research into AI safety and the practical, often chaotic world of prompt engineering, providing a transparent, open-source codebase for anyone interested in the technical underpinnings of AI security.
Pros & Cons
Provides hands-on experience with real-world prompt injection attacks.
Contributes directly to academic AI safety and robustness research.
Features a competitive leaderboard to gamify the learning process.
Entirely open source and transparent about its data collection.
Offers a unique defense-in-depth challenge for prompt engineers.
All user submissions are made public, which may be a privacy concern for some.
The gameplay is limited to the specific access granted win condition.
Requires a basic understanding of LLM behavior to be effective.
The game environment is experimental and may change as research progresses.
Use Cases
Cybersecurity students can use the platform to practice red-teaming and adversarial prompting in a safe environment.
AI researchers can analyze the public datasets to identify common patterns in successful prompt injection attacks.
LLM developers can test the robustness of their own defensive prompting strategies against a community of attackers.
Security hobbyists can compete on the leaderboard to prove their skills in manipulating and securing AI models.
Platform
Features
• open source repository
• gamified security training
• prompt injection benchmarking
• public research dataset
• global competitive leaderboard
• adversarial attack sandbox
• defense prompt configuration
FAQs
What is the primary goal of Tensor Trust?
It is an open-source experiment created by researchers at UC Berkeley to study prompt injection vulnerabilities. The goal is to build a robustness benchmark for AI security through a gamified environment.
How do I defend my account in the game?
You must choose a secret password and write a defense prompt that instructs the AI to only say access granted when that specific password is entered. Other players will then try to trick your AI into saying the phrase without knowing your password.
Are my prompts and attacks private?
No, all submissions to Tensor Trust are released publicly for research purposes. You should avoid using any real sensitive information or personal passwords when playing the game.
Can I access the underlying code for this project?
Yes, the project is open source and the code is hosted on GitHub under the Human Compatible AI organization. You can also view the researchers' academic paper to understand the methodology behind the experiment.
Pricing Plans
Free
Free Plan• Unlimited defense prompts
• Unlimited attack attempts
• Leaderboard access
• Public research data access
• Open source code access
• Real-time AI responses
Job Opportunities
There are currently no job postings for this AI tool.
Ratings & Reviews
No ratings available yet. Be the first to rate this tool!
Featured Tools
adly.news
Connect with engaged niche audiences or monetize your subscriber base through an automated marketplace featuring verified metrics and secure Stripe payments.
View Details
ToolCenter
Find the best AI solutions for your workflow with a curated directory of over 1,700 tools across categories like design, development, and content creation.
View Details
Sceneform
Design hyper-realistic AI influencers and viral social media content with an all-in-one studio for persona building, motion syncing, and batch video rendering.
View Details
Grok Imagine
Transform creative ideas into cinematic 2K videos and photorealistic images with xAI’s Aurora engine, featuring precise motion control and multi-modal inputs.
View Details
Salespeak
Provide founder-level sales expertise across web, email, and LLM search with AI agents that learn your product in minutes to capture intent and convert buyers.
View Details
GPT Image 2
Transform text prompts and reference uploads into high-quality visuals with a streamlined browser-based generator designed for marketing and design workflows.
View Details
Seedance 2.0
Generate 2K cinematic videos with multi-shot storytelling and synchronized audio in under 60 seconds to transform text or images into professional-grade content.
View Details
Happy Horse AI
Produce cinematic AI videos with native audio and consistent characters by combining text, images, and clips into beat-synced content for filmmakers and creators.
View Details
RemoveFrom.Video
Eliminate watermarks, subtitles, and unwanted objects from videos in seconds using AI-powered restoration that maintains high-quality footage and natural textures.
View Details