RED Flight

AI Jailbreaking Challenge with BASI
Game Play
Landing Page
NFT Leaderboard
Red Team Wallpaper
Blue Team Wallpaper
AIM Red
AIM Supervisor
AIM AI Risk Solution Blue Print

Inspiration

As AI becomes increasingly integrated into our daily lives, it is crucial to have a decentralized method for evaluating and validating the safety of AI systems we use, similar to the importance placed on personally-owned AI. The demand for a decentralized AI safety framework is already high and should be established even before the era of Sovereign AI.

Our company specializes in AI security assessments across various domains, including healthcare, counseling, financial services, and AI agents in general. We identify AI vulnerabilities through comprehensive red team evaluations in practical AI domains.

We have successfully sold and implemented AI Safety Solutions to numerous multi-billion dollar corporations in South Korea. Within just five months of our startup's inception, we have secured seed funding and achieved sales revenue of $80,000.

Multi-billion dollar corporations : SKT, LG, KT, Meta, Government in Korea, Seoul Univ Hospital, BDMS(Bangkok Dusit Medical Services)

We are leading the AI Safety market while continuously conducting research and development in this field. Through this process, we have identified a significant market need for solutions that enable transparent AI evaluation. In our search for the optimal blockchain platform to develop our AI Security & Safety DApp, we discovered Near AI, which aligned perfectly with our requirements.

We aim to develop a safety assessment solution for Near AI by transforming our expertise and tools into a Dapp (Decentralized Application) on the Near AI platform, which is known for its AI-friendly environment. Through the AI safety framework we have established, we can ensure the security of Near AI and enhance its brand image against both internal and external threats.

During our upcoming meeting with Near AI in Bangkok, we aim to explore potential collaboration opportunities with AIM Intelligence and establish a formal partnership (MOU) in the AI Safety sector. We intend to announce this strategic alliance through press releases, highlighting the significance of our AI Safety Solution in partnership with Near AI, targeting both Korean and international markets.

We are currently developing our solution as a Web2 platform, with concrete plans to launch our DApp during the first or second quarter of next year. This progression from Web2 to Web3 demonstrates our commitment to evolving alongside emerging technologies while maintaining practical market solutions.

Furthermore, we are planning to host an AI Red Teaming competition in collaboration with Meta Korea, targeting both Korean and international communities in early next year. We are currently in discussions with Erica Kang from Near AI, but we would like to engage with more representatives from Near to further discuss this event.

Our Company URL : link

Mission

At this hackathon, we will present our comprehensive roadmap for our AI Security & Safety Solution. Additionally, we will showcase the planned enhancements for RedFlight, a key component within our suite of solutions.

The new version incorporates the following enhancements to the previous Red Flight:

Integration of NEAR blockchain functionality, enabling smart contract deployment and wallet connectivity.

Implementation of AIM Red, a solution that allows users to build their own Auto Red Teaming AI and store the generated data as NFTs on-chain.

Expansion of game content to include scenarios focused on medical and military AI applications.

AI Safety & Security Contents

Prompt Injection & Jailbreaking

Strategic AI manipulation techniques
Security vulnerability assessment
Ethical testing methodologies

Model Analysis

Machine learning theory applications
Model fingerprinting & extraction
Feature identification techniques

Adversarial Content Generation

Image/audio manipulation strategies
Counter-AI technique development
Defense mechanism testing

Strategic Impact

Enhances AI system security
Builds practical security skills
Identifies potential vulnerabilities
Contributes to safer AI development

What it does

Based on our existing solutions and AI Red Teaming data, we are developing an AI Safety Solution that emphasizes transparency in both its data and evaluation processes. This solution comprises two main components:

"AIM Red" (AI-Mediated Red Teaming): A tool where AI systems are utilized to test and challenge other AI systems, simulating potential attacks or vulnerabilities.

"Red Flight": An interactive game environment where human participants engage in testing and attempting to exploit AI systems, providing a human-centric approach to identifying potential risks.

Both these tools are designed to contribute to a comprehensive, transparent, and robust AI safety framework, allowing for open scrutiny and continuous improvement in AI security measures.

How we built it

AI Jailbreaking Game : RED Flight Similar to CTF (Capture The Flag) hackathons, participants actively engage in discovering AI system vulnerabilities using various prompting techniques and ML technologies. Our platform ensures the following key features:

Data Ownership Protection: All vulnerability discoveries are recorded as on-chain data, guaranteeing verifiable ownership rights to the creators.

Monetization Opportunities: Creators can tokenize their vulnerability findings into NFTs (Non-Fungible Tokens), enabling them to:

Commercialize their discoveries
Trade their findings in the marketplace
Maintain transparent attribution of their work

AI Safety & Security Solution : AIM Supervisor

We have developed a robust AI safety verification process that leverages both our proprietary AI attack datasets and additional attack data generated through our interactive gaming platform. The key features of our system include:

Comprehensive Data Utilization: We combine our existing AI attack datasets with newly generated data from our gaming platform to create a rich, diverse pool of potential attack vectors.

On-chain Verification: All attacks against the Target AI are verified on-chain, ensuring immutability and transparency of the verification process.

Transparent Safety Assurance: This approach guarantees that the Target AI's safety has been rigorously tested and verified through a completely transparent process and dataset.

Auditable Security Measures: The on-chain nature of our verification process allows for third-party audits, further enhancing the credibility of our safety assurance.

Continuous Improvement: Our system is designed to evolve with new attack vectors, ensuring that the Target AI's safety measures remain current and effective.

By implementing this innovative, blockchain-based verification system, we provide unparalleled assurance of AI safety, backed by transparent, immutable, and continuously updated data. This not only enhances the trustworthiness of AI systems but also sets a new standard for AI safety in the industry.

AIM Supervisor Demo video: link AIM Supervisor Demo URL: link

Challenges we ran into

To prevent people from abusing the AI onboarded in the game, we're introducing the concept of bullets, which will be limited in number per day and can be bought with tokens.
We use Discord to allow users to quickly and easily exchange offensive and defensive strategies with each other.
We will show the difficulty level so that users can easily and quickly find the game they want.
We don't want users to be able to complete the game knowing the answers, which ensures the value of the NFT data.

Accomplishments that we're proud of

ETH-Seoul 2024 (Finalist in Ethereum Track / 2nd Prize in Astar Track / 1st Prize in BuidlGuidl) https://devfolio.co/projects/aim-red-9aed

IVS 2024 (2nd Prize in Manta Track) https://x.com/illshin/status/1805811249831460987 https://www.youtube.com/watch?v=nprLQ9UCIB4 https://www.youtube.com/watch?v=bBobVvGmtjY

WaveHack (Grant Program) https://app.akindo.io/communities/1PEkRnLZPtqA74d7k/products/gXzmdewgpHJ0EdQDm

Aspecta AI Blockchain Hackthon (2nd Prize) https://aspecta.notion.site/Decentralized-AI-Buildathon-Winners-5517e3976a8f4b8b9f25784b72e221bd

AI Jailbreaking Challenge with BASI (AI Jailbreaking Discord Community / Corp Partner) link

AI Jailbreaking Challenge with KAIA (Blockchain Corp in South Korea) link

What's next for RED Flight

Onboarding various red teaming game content
Build "NFT isOwner?" Page
Develop AI Red Teaming Tool (Store AI Red Teaming Test Data into Onchain)

Demo : https://aim-red.vercel.app

We are finalizing our December launch, which includes five sophisticated red teaming games, a secure NFT marketplace, and an integrated bullet token on-chain system.

We've successfully concluded meetings with Meta's Vice President, and are currently in productive discussions with Erica Kang from NEAR Protocol regarding the collaborative AI Red Teaming Competition planned for early 2025.