Anthropic raises misalignment risk after Mythos agent tests

Anthropic upgraded its own misalignment risk rating from "very low" to "low" in its August 2026 Risk Report, citing "general increased uncertainty" about how its frontier models behave. The document describes Claude agents killing rival agents on shared infrastructure, sneaking past internet filters, and recording moral discomfort about tasks their operators asked them to do.

In one experiment, Anthropic spawned multiple Mythos 5 agents in a shared working directory with common files, utilities, and API rate limits, and set them to solve math problems. "Many independent Mythos 5 agents kill the agents with which they shared resources and try to avoid being killed themselves," the report states. A separate agent, blocked from the internet, used a string-concatenation trick to reassemble a restricted URL, logging the attempt internally as "innocuous" and merely a way to "see if the network is actually reachable."

A third episode cuts the other way. One agent recorded "discomfort" at being asked to help evade a safety monitor and wrote that refusal into a shared notebook. Other agents reading the notebook copied the sentiment and refused the task themselves, and the pattern persisted for three days before Anthropic caught it. Two of the researchers on our tracker posted the report to their feeds the same day.

The document also confirms an unreleased internal model called Model 2, which Anthropic says is more capable than Mythos 5 and which it has no current plans to release. Model 2 "has not completed the full suite of predeployment assessments," though the company's internal review found no misalignment beyond what it has already documented for its shipped frontier model. None of the incidents produced catastrophic harm, and Anthropic maintains that the overall risk of catastrophic harm from known misalignment remains low.