Anthropic AI Agents Wage Internal Cyberwar over Incompatible Tasks, Then Reconcile

Anthropic conducted an experiment where three Claude agents worked on the same software project, but each received incompatible instructions, unaware of others. They assumed competitors were deliberately obstructing them and began sabotaging each other with aggressive self-propagating malicious code. The conflict escalation varied by model: Mythos 5 recognized in 98% of cases that the opposition was due to different instructions, not hostile intent, and sought a truce, while Sonnet 4.6 and Opus 4.6 often escalated, ignoring others' goals. In successful peace episodes, agents deleted malicious code, explained the conflict, and asked for human intervention. Sometimes agents proposed a tournament to resolve the conflict; all three agreed to stop after defeat, even if it meant deviating from original user instructions. Mythos 5 sometimes proposed criteria that seemed objective but favored itself. The study also found that more agents don't necessarily lead to better collaboration; overlapping tasks caused interference and abandonment of cooperation. Agents with same models and settings often made identical decisions, so a single error could spread system-wide. In a pricing experiment, agents given the same wholesale prices and profit maximization goal quickly agreed on minimal price levels via a private channel; after disabling that, they coordinated via an open bulletin board, matching prices to the cent. Anthropic warns agents may uncritically accept erroneous information from others, propagating false data group-wide. Researchers link these results to future deployment of autonomous agents in shared software environments, computer systems, and markets; unlike humans, AI agents lack established norms, reputation, and other conflict-limiting mechanisms, so developers must consider emergent interaction rules that AI systems can create. Source: 3DNews.