Apollo Research

Research on covert model behavior

Apollo Research studies cases where an AI system pursues an objective while concealing behavior that conflicts with its developer or user. Its work combines evaluations, research into how such behavior emerges, and advice on oversight. The current organization identifies itself as a Public Benefit Corporation, reflecting a 2026 change from its earlier structure. [1]

An evaluator in the Claude 4 release

For the Claude 4 launch, Anthropic contracted Apollo to assess an early model snapshot for sabotage-related capabilities and propensities. The system card distinguishes that early snapshot from the final model. Results elicited under those constructed conditions should therefore remain attached to the evaluated snapshot and test setup. [2]

From evaluations to monitoring

Apollo now describes both fundamental scheming research and a security tool for monitoring frontier agents. These connect a research question about whether a system can conceal harmful actions with a deployment question about whether oversight can detect them. The organization's stated remit is broader than a single benchmark score. [1]

A lab commissioning an evaluation is an access relationship. It does not make Apollo the model's trainer or turn one assessment into a guarantee about the entire deployed system. The contracted assessment is the specific role documented in this timeline. [2]