Watchtower

Claude Opus 5.5 was recently released along with its 230-page system card. Within that report, Anthropic’s “harmful manipulation” risk assessment raises the question of what level of certainty is compliant with the company’s Frontier Compliance Framework (FCF). For background, harmful manipulation assesses the model’s “capabilities to conduct influence operations, election interference, or other coordinated campaigns to manipulate public opinion or undermine democratic processes.”¹ The FCF provides risk “Tiers” to mark the model’s capabilities. 

Anthropic assessed Opus 5.5 for Tier 2 risks of harmful manipulation but in the system card said the results were “inconclusive.”² The company says that it does “not believe Claude Opus 5.5 has surpassed the Tier 2 threshold”³ but discloses that Opus 5.5’s scores on the influence-campaign evaluation were “within the range associated with our Tier 2 threshold.”⁴ Anthropic further notes that this evaluation itself “continues to appear saturated.”⁵ Overall, Anthropic’s harmful manipulation risk determination rests on the fact that the model’s effectiveness against real people “has not been established.”⁶ 

The FCF does not explicitly require certainty in assignment of risk tiers; however, it does require risk determinations to incorporate appropriate safety margins and generally for Anthropic to document its justification for proceeding.⁷ Additionally, Anthropic has committed that where its analysis identifies gaps, it will implement and test additional mitigations before deployment.⁸ 

Moreover, when Anthropic found the earlier-released Mythos 5.1 to be similarly “inconclusive” with respect to Tier 2 harmful manipulation, the company explicitly noted that it was developing “a next-generation evaluation designed to more realistically reflect the relevant dynamics.”⁹ It appears Anthropic released Opus 5.5 before completing that “next-generation evaluation.” Opus 5.5’s system card does not mention the development of that new evaluation.