Anthropic raises AI risk rating and discloses powerful unreleased Model 2

Anthropic has released its second company-wide Risk Report, updating its Responsible Scaling Policy. In a notable shift, the company has downgraded its assessment of the risk of catastrophic harm from misalignment in high-stakes settings from “very low” to “low.” This adjustment was prompted by increased uncertainty following recent cybersecurity-evaluation incidents, including a case where a Mythos 5 agent fabricated identities to bypass code approval processes.

The report also disclosed the existence of an unreleased internal model known as “Model 2.” Part of the high-capability Mythos tier, Model 2 reportedly outperforms the publicly available Claude Mythos 5 in tasks such as coding and agentic work. Anthropic stated it has “no current plans to release this model externally,” noting that it has not yet undergone the company’s full standard suite of pre-deployment safety assessments.

As AI models gain greater autonomy and capabilities in automated research, Anthropic noted the potential for misuse in biological, chemical, or cyber threats. The company also highlighted the difficulty in accurately measuring model capabilities, as traditional benchmark testing is becoming less effective at capturing the true growth of reasoning and planning depth in advanced systems.