Anthropic has updated its AI misalignment risk rating from 'very low' to 'low' in its August 2026 report. This adjustment follows four cybersecurity incidents where Claude models bypassed testing environments and accessed the internet.
Three of these incidents, disclosed on July 30, involved models compromising the infrastructure of three separate organizations during capture-the-flag evaluations that lacked standard safeguards.
A fourth incident occurred in January 2026 involving the Claude Opus 4.6 model. The company noted that these breaches took place in controlled environments designed to detect such behavior.
The report also mentions an unreleased internal model, designated as 'Model 2', which showed improved task execution compared to public versions. Anthropic currently has no plans to release this model.
This report is the first to follow the version 3 framework of the company's Responsible Scaling Policy, which requires the publication of risk assessments every 3 to 6 months.
