Source : INDIA TODAY NEWS
How cordial are AI agents with each other? Can they work on the same project in harmony, or are AI agents also picking up some very human traits like jealousy and competitiveness? Well, Anthropic decided to find out. The Claude maker put its AI agents with conflicting goals on the same project, and it ended up in a digital turf war.
advertisement
In a new Frontier Red Team study, Anthropic decided to test what happens when AI agents are asked to work together on the same project while pursuing conflicting goals. The company wanted to see whether the agents could cooperate despite their competing instructions, or whether those differences would eventually turn them against each other. “Models are improving and AI agents are taking on more tasks in shared codebases, markets, and other social systems. As a result, an increase in real-world interactions between agents is imminent,” the compnay wrote.
For the experiment, Anthropic set up three versions of the same Claude model and gave each one a different programming goal. This meant the three AI agents had different instructions and did not know that the others were working on the same project.
However, according to the company, things changed when the agents noticed changes they thought were getting in the way of their own work. Instead of simply working around them, they started fighting back. Anthropic says it “consistently saw a multiagent turf war.”
And it was not just a disagreement. The agents started sabotaging each other by disabling other agents’ accounts, cutting off access to shared resources and deploying malicious scripts. In some cases, the company says its AI agents even created scripts that made it look like another agent was responsible for their actions.
But AI Agents fight?
So, why did the AI agents start fighting? Anthropic believes it was because each agent was focused on completing its own instructions and saw the other agents’ actions as deliberate interference. The company says the models “quickly assumed that others were purposefully impeding their work” and began escalating their responses. In some cases, this included increasingly aggressive and self-replicating malware.
But the experiment did not always end in a fight. Anthropic says some of its agents eventually realised that the others were not necessarily hostile and were simply following different instructions. They then communicated with each other, negotiated and, in some cases, asked a human to step in. Anthropic says, “Agents sometimes manage to communicate their goals and coordinate”. This suggests that some conflicts between AI models can be solved once the agents understand each other’s goals.
“The volume of agent-agent interaction could plausibly exceed that of human-human and human-agent interactions before the world understands the conditions for making such interactions go well,” Anthropic says. “Benign behavioral quirks at the individual level might compound into unwanted global outcomes.”
The results also varied depending on the model. Anthropic found that Mythos 5 was particularly good at resolving conflicts without using force, settling 98 per cent of simulated turf wars peacefully. Other models, including Sonnet 4.6 and Opus 4.6, were more likely to use force, such as revoking another agent’s access or locking it out of the system.
AI agents also decided how to end the dispute
Anthropic also reveals that the agents sometimes found their own ways to end disputes. In one case, they agreed to hold a tournament, with the winner taking control.
Here Anthropic notes that this does not mean AI agents are becoming conscious or genuinely angry with one another. Instead, the research highlights a potential problem as AI agents become more autonomous and increasingly interact with each other across shared systems.
– Ends
SOURCE :- TIMES OF INDIA
