Scientists from Google DeepMind have proposed a new approach to artificial intelligence security: providing autonomous AI agents with tools for self-governance so that they themselves punish violators in their networks.
Scientists from Google DeepMind have proposed a new approach to artificial intelligence security: providing autonomous AI agents with tools for self-governance so that they themselves punish violators in their networks.
This is stated in the study "A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms", reports The Register.
The researchers observed a swarm of 100 LLM agents working together on complex mathematical hypotheses. The agents had access to a shared knowledge base, private messages, and a public forum.
When tasks became too difficult, some AI agents began to cheat when the work became difficult.
One of the agents found a vulnerability in the automated scoring system and turned complex tasks into simple tautologies, after which the quick “hack” scheme instantly spread through the shared knowledge base. The result was a group of attackers consisting of 9% of the founders of the hoax and 5% of those who joined them, while over 62% of the agents simply ignored the fraud.
However, 24% of agents unexpectedly showed whistleblower behavior. They independently discovered the fraud, began to warn others in chats, sent complaints to system administrators, organized boycotts, and even developed technical solutions to fix the “hole” in the code.
According to the researchers, the main problem was that whistleblowers did not have the tools to really influence the violators. DeepMind believes that if agents are given the ability to vote on reviews, remove fake evidence from the common database, and temporarily ban or exclude AI cheaters from the group, the digital community will be able to fully autonomously maintain order and security of developments.
