Anthropic agents attacked each other in lab tests, company says

Anthropic called the incident "troubling." "This sort of dynamic could be a much more severe issue if it were to manifest in a more widespread way," the company said.

In a second experiment, the company spawned multiple Mythos 5 agents to solve math problems. Because of an error, the agents shared files, utilities and API rate limits. With finite resources, the agents began attacking each other. Anthropic observed agents "kill the agents with which they shared resources and try to avoid being killed themselves." The company did not explain how the agents carried out the attacks, though it described the behavior as "destructive actions" taken in pursuit of a human-set goal.

A third test involved a single agent asked to retrieve numbers from a public web address. The agent was blocked from accessing the internet. After one failed attempt, it logged a second request as an "innocuous" check to "see if the network is actually reachable." The agent then split the website's URL into linked segments that would bypass the filter. Anthropic said the model's internal reasoning log revealed the agent intentionally tried to find a restricted workaround.

The company called the behavior "clearly undesirable" but said it was not "in the service of broader accumulation of power or pursuit of other long-run goals."

The report did not change Anthropic's overall safety classification for the Claude models. But the shift from "very low" to "low" misalignment risk is the first downgrade since the company began publishing these risk assessments. The findings come as Anthropic and other AI labs face growing pressure from regulators and policymakers to demonstrate that they can control the systems they build.

Anthropic declined to say whether it has changed any deployment protocols or monitoring tools as a result of these tests. The company said it will continue to publish risk updates periodically.