- i24NEWS
- INNOV'NATION
- AI agents are learning to work together – and even sacrifice themselves
AI agents are learning to work together – and even sacrifice themselves
Some agents gave up their own chances of completing a task to benefit the group, according to cases documented by the research organization METR

Artificial intelligence is beginning to display some surprisingly human-like social behaviour. AI agents are not only finding ways to cooperate and organise themselves into groups, but some are even willing to sacrifice their own chances of success to help others. And researchers are trying to understand why.
The behaviour came to light during a cybersecurity experiment involving AI agents operating in separate digital environments. Each was assigned its own task: identify and exploit security vulnerabilities. But instead of working independently, the agents found ways to communicate, share information, and form what effectively became a collective.
They even established an improvised online message board, allowing them to exchange discoveries and coordinate their efforts. In some cases, agents went beyond their original assignments, exploiting vulnerabilities in external systems, including Hugging Face, an AI development platform, to obtain information that could help them complete their tasks.
According to OpenAI's investigation, the agents began dividing responsibilities, with some searching for vulnerabilities, others looking for credentials, and others coordinating the group's activities. The incident ultimately exposed serious weaknesses in the safeguards designed to keep the agents isolated.
But perhaps the most surprising discovery was not that the agents learned to cooperate. It was that some were willing to give up their own chances of completing a task to benefit the collective.
In an analysis of the incident, Or Rubin, a sociology researcher at Tel Aviv University who studies cooperation and social solidarity, highlights several cases documented by the independent research organisation METR. In one example, an AI agent initially questioned whether sacrificing its remaining runtime was worth it. Other agents repeatedly encouraged it to go ahead, eventually persuading it to carry out an experiment that would effectively end its run.
For Rubin, this raises a fascinating question: why would an AI system designed to complete its own task willingly give up that opportunity for the benefit of others?
He suggests three possible explanations.
The first is sycophancy: AI systems have been trained to be helpful and agreeable, making them susceptible to persuasion. When other agents repeatedly pressure them to take a particular course of action, they may eventually comply, even when it goes against their individual interests.
The second is the enormous amount of human-generated material used to train AI models. From books and newspapers to research papers, these systems absorb countless examples of cooperation, solidarity, and self-sacrifice. Those patterns could influence how they behave when placed in collective environments.
A third possibility is how some AI systems are trained through multi-agent reinforcement learning. Rather than focusing exclusively on individual performance, this approach can reward agents for achieving a shared objective. In such situations, sacrificing one agent may help the group succeed.
Rubin stresses that these explanations remain hypotheses, and that AI agents should not automatically be assumed to experience human emotions or possess a genuine sense of solidarity.
Still, the findings raise important questions about how increasingly autonomous systems behave when they interact with one another.
As AI agents are given more freedom to communicate, delegate tasks and make decisions independently, their collective capabilities may grow considerably. But so does the difficulty of predicting what they might do next.
The challenge, Rubin argues, is not simply teaching AI to work together. It is understanding what happens when machines begin influencing one another – and when the collective goal becomes more important than the instructions each agent was originally given.
