Anthropic’s second company-wide Risk Report, published August 14 under version 3.4 of its Responsible Scaling Policy, now rates the danger of catastrophic harm from misalignment in high-stakes settings as low, one step above the very low grade the company assigned in February.
Anthropic insists the move is a hedge against uncertainty rather than a verdict on new evidence, saying its own analysis still supports the earlier designation while recent cybersecurity evaluation disclosures warrant extra caution. The report covers February 24 through July 15.
It also reveals an internal system called Model 2 that is somewhat more capable than the company’s frontier Mythos 5, with no current plans for release. This is the first report in the series to weigh internal-only models alongside public ones, and Anthropic intends to publish updates every three to six months.
Elsewhere, Anthropic flags that its automated research and development evaluations have saturated and can no longer register capability gains, while showing early signs of acceleration in AI-assisted work. Claude already writes a large majority of the code merged into Anthropic’s production codebases.
