It also said there was a similarly low risk of highly capable AI being able to "perform automated research and development" which could cause "catastrophic harm initiated by the AI". But it said it was "less confident in this assessment" than it was previously.