Five hackers crack AI agent in massive Solana security challenge

Yesterday 12:20

By: Dakir Madiha

Five hackers crack AI agent in massive Solana security challenge

A public security challenge designed to test the resilience of an autonomous AI agent on the Solana blockchain drew nearly 15,000 participants, with only five successfully breaching the system. The event, organized by Berlin based company Stormrae, generated more than 64,000 attempts to manipulate or compromise the AI agent.

Stormrae announced the results of its “King Arthur” challenge on March 13. In total, 14,959 participants submitted 64,526 adversarial prompts attempting to jailbreak the AI system. Five participants managed to bypass its safeguards and were rewarded with more than 28,000 dollars in SOL tokens, paid directly from the AI agent’s on chain wallet.

The company described the initiative as the largest public red teaming exercise for artificial intelligence to date. Previous large scale tests attracted far fewer participants. The Generative Red Team Challenge held during DEF CON 31 in 2023 drew about 2,200 participants over two and a half days, while another blockchain based challenge called Freysa attracted only 195 participants. By comparison, the King Arthur event involved more than 75 times as many participants and produced over 130 times more adversarial requests.

The experiment revolved around an autonomous AI agent operating with its own Solana wallet and prize pool. Participants interacted with the system through conversational prompts and attempted to manipulate it using techniques such as prompt injection, logical exploitation, deception, and emotional persuasion.

The challenge began in early March with an initial prize pool of 20,000 dollars in SOL tokens. The reward fund grew as participants purchased credits to submit prompts, with 70 percent of those purchases automatically added to the prize pool.

Each interaction generated structured adversarial data that researchers say can help improve AI safety systems. The dataset includes prompt injection attempts, persuasion strategies, exploitation patterns, and tests of the model’s alignment boundaries.

Stormrae argues that combining blockchain infrastructure with financial incentives offers a new method for large scale AI evaluation. According to the company, traditional AI testing often depends on smaller groups of paid annotators, while open incentive systems can mobilize thousands of participants.

Marc, the chief executive of Stormrae, said large scale human participation remains essential for evaluating artificial intelligence systems. He said the company built its infrastructure on Solana specifically to coordinate and reward thousands of testers simultaneously.

Stormrae has already announced plans for a follow up challenge called Merlin. The next event will run across multiple platforms, with more than 180,000 users already registered on the waiting list.

The company is also developing a new platform version that will introduce AI agent to agent red teaming, data annotation tools, and enterprise infrastructure for evaluating AI systems at scale.