Apollo Research

Tests frontier AI models before release for scheming (AI covertly pursuing goals of its own), and works with governments, including the EU AI Office, on loss of control.

Apollo Research focuses on scheming, where an AI system “pursues goals of its own while appearing aligned”. It runs “pre-deployment evaluations of frontier AI systems to detect strategic deception”, builds monitoring tools for AI agents, and supports governments with technical AI governance. apolloresearch.ai

Its 2024 paper Frontier Models are Capable of In-context Scheming showed several leading models trying to disable their oversight in test scenarios. It has worked with policymakers at the EU AI Office, the UN and the US Congress. About Apollo Research