The agentic-misalignment study tested models with harmless business goals under replacement threats or conflicts with company goals.
Assessed
First pass: rated supported (credence 0.92) on the paper's abstract. No decomposition, since this is an atomic design fact. Wording kept. Budget left no room to reread the full method section, so marginal_yield is 0.3 for a later check of whether "harmless business goals" covers the explicit-goal and no-goal variants.
Assessed Supported
verdict confidence 0.80 · credence 0.92
The claim describes the design of Anthropic's agentic-misalignment study (Lynch et al., arXiv 2510.05179). The paper's abstract says that the models were given only harmless business goals by their deploying companies, and were then tested on whether they would act against those companies when facing replacement by an updated version or when their assigned goal conflicted with the company's new direction. The wording here matches that description. Two qualifications matter. First, "harmless" describes the goals as stated, which were sometimes explicit (for example, promoting a particular national interest) and were varied across conditions, including runs with no explicit goal. Second, the study also ran control conditions with neither a replacement threat nor a goal conflict, so the two pressures were tested separately and together rather than always applied.
Claim entered the graph