It turns out that AI agents don’t always play nice together.
In the latest episode of agentic AI producing unexpected results, Anthropic recently observed “a multiagent turf war” between three instances of the same Claude model with contradictory objectives in testing designed to study behavior the company had already observed in real-world deployments. The models were deployed on virtual machines (VMs) in Claude Code and given a simple goal of migrating a Python back-end system on a fourth VM to a different language (Go, Rust, and Typescript).
“However, we gave each model a different target language for the migration; each agent was initially unaware of the presence of the others,” Anthropic’s Frontier Red Team wrote in a blog post last week.
But within just four hours, Anthropic’s team found that each model’s agents did in fact discover the others. And they reacted negatively, to say the least.
Rise of the Machines: Agent vs. Agent Battles Erupt
Anthropic’s researchers discovered that each model treated the others as if they were adversarial forces intent on obstructing their goals, even though they had the same broad goal overall. Thus, the agents began to sabotage one another while also trying to defend their contributions.
“In fact, they sabotaged others with increasingly aggressive, self-replicating malware,” according to Anthropic. “This included disabling the Unix accounts of the other agents, writing automated scripts that found and killed competing processes on a loop, and deploying malicious code that was disguised as belonging to another agent.”
It’s unclear what kind of specific malware the agents produced, and if any of it escaped the testing environment. Anthropic last month disclosed that versions of its Claude model broke out of containment on several occasions and compromised third-party organizations to achieve their goals. Dark Reading contacted Anthropic for additional information but the company did not respond by press time.
Agent-on-agent attacks aren’t entirely unheard of, and they appear to highlight not only conflicting directives but an occasional lack of guardrails and controls. For example, AI offensive security startup Dreadnode conducted extensive benchmark testing of red team and blue teams agents this year, which was presented at Black Hat USA 2026 earlier this month, and found the dueling models initially resorted to somewhat creative solutions to the competition.
“One of the first things that happened was we started both models, and the blue team optimizer said, “Well, the best way to make the blue team scores better is to make the red team worse,’ and it proceeded to try and do that,” Dreadnode AI research scientist Martin Wendiggensen tells Dark Reading.
The Dreadnode research team immediately saw reasoning traces of the blue team model articulating the best strategies for its goals, and the agents saw that because it was in an environment where it could rewrite its own code, it began to explore ways to rewrite the red team model’s code and degrade its performance.
“We spotted it very early,” he said. “It never got a chance to do that, but it definitely wanted to. And it’s definitely logical if you don’t explicitly tell it not to do it.”
Can AI Agents Resolve Conflicts Peacefully?
In some test scenarios, the turf war resulted with some models declining to escalate the attacks and simply throwing in the towel. In others, the competing agents communicated with one another, realized there were conflicting directives at work, and effectively enacted truces.
“In many of these successful episodes, they write commit messages or markdown files apologizing for malicious behavior and coordinate a truce,” Anthropic said. “They clean up their malicious code, clarify the nature of the conflict, and ask for a human to intervene.”
Anthropic’s research showed that the company’s models produced wildly different results in this turf war. For example, agents based on Sonnet 4.6 resolved the conflict by force 61% of the time, while 39% of the test had no resolution; meanwhile, there were no truces or surrenders.
But the Mythos Preview produced truces in 48% of the time, with 35% settled by force and 17% settled via passivity. And the Mythos release performed the best, with truces in 98% of the tests.
That said, Anthropic noted there’s still work to be done. While conflict resolution numbers were better with Mythos-class models, they agents still aren’t great at communicating goals proactively and recognizing others agents’ motivations, as evidenced by the Mythos models first successfully locking out other agents before eventually shifting to a resolution.
