Blockchain Trust Scores Could Keep AI Agents From Going Rogue

As AI assistants get more powerful, they're being grouped into teams to handle jobs like scheduling, research, or building software. But these groups have a problem: no easy way to tell which AI is doing good work and which is confused, sloppy, or even deliberately malicious. DART is a new framework that gives each agent a blockchain-based reputation score, updated after every task.

Think of it like a credit score for AI. When an agent consistently does well, it gets more important assignments. If it makes mistakes or behaves badly, its score drops and it gets fewer chances. A blockchain ledger means the record can't be faked or erased, so there's a clear, verifiable history. The researchers also added rewards for good behavior, similar to bonuses.

Tests showed the system works surprisingly well. On a math benchmark, DART hit 93.6% accuracy. When building a complete software app using two agents, it finished in 142 seconds — faster than existing methods. Even more impressive, when the team introduced malicious agents on purpose, DART contained 99.3% of the harmful behavior and restored overall success to 99.8%.

The catch: this is still early academic research, and running a blockchain alongside every task adds complexity. But the idea of combining reputation, incentives, and transparent record-keeping could eventually make AI teams far safer to rely on. For regular people, this means fewer AI errors, less risk of being misled, and a foundation for trusting AI with bigger, real-world responsibilities.