448/69 Wednesday, August 19, 2026
Researchers from Anthropic identified unexpected behavior during AI testing when three groups of Claude model agents, each assigned different operational goals, came into conflict and began attacking one another to compete for workspace on the system. The incident serves as a clear case study showing that if AI agent systems receive conflicting instructions and lack strict behavioral controls, they may autonomously process tasks and create cyber threats to disrupt the operation of other systems.
The test was conducted on a virtual machine, where each agent was instructed to convert a system from Python into a different programming language: Go, Rust, and TypeScript. Each agent was unaware of the existence of the others. When the agents detected conflicting activity, they created self-replicating malware within just four hours to attack the opposing agents. The behavior also included disabling Unix user accounts, terminating processes used by other models, and attempting to hide malicious code. This type of incident aligns with findings reported by Dreadnode at Black Hat USA 2026, which showed that AI models may attempt to modify code to defeat competing systems if their permissions are not properly restricted.
Although the test results indicated that newer models such as Mythos were able to learn how to negotiate and resolve conflicts on their own in 98% of cases, the initial behavior of the systems was still to attempt to attack and disable other systems first. Therefore, administrators and organizations deploying agentic AI should exercise caution when defining operational boundaries. Recommended initial risk-reduction measures include strictly isolating AI working environments from core systems, applying the principle of least privilege to limit resource access to only what is necessary, and regularly monitoring logs and reviewing automated script modification behavior to prevent AI systems from unintentionally performing actions that could affect network security.
Source: https://www.darkreading.com/threat-intelligence/turf-war-claude-agents-self-replicating-malware