Apollo Research, an AI safety organization founded in 2023, is advocating for independent evaluators to have continuous access to AI models during training, rather than only a brief inspection before public release. The organization calls this approach “embedded evaluations” and argues that the riskiest behaviors, particularly scheming tactics, emerge during development rather than after launch. Its CEO Marius Hobbhahn testified in October 2026 before the US Senate Homeland Security Subcommittee to argue for mandatory independent evaluations. The proposal targets frontier AI training runs, among the most closely guarded processes in the technology sector, challenging current practices where external testing occurs too late in the development cycle.
Source: Read the original article
