Apollo Research

Ce qu'il fait

Evaluates frontier models for scheming, strategic deception and evaluation-awareness; builds Watcher, a runtime monitoring tool for coding agents; and advises governments on frontier regulation. Ran the in-context scheming evaluations that appear in OpenAI and Anthropic system cards.

Évaluation franche

Small, focused, and explicitly soliciting research collaboration on its contact page — one of the more genuinely open technical labs. There is no dedicated route for reporting a concern about a model or a company, only a general inbox.

Préoccupations couvertes

Scheming — AI that pursues its own goals while appearing aligned — plus loss of control in automated AI R&D and misaligned models as insider threats. Explicitly a deceptive-alignment organisation.

Canal gouvernemental

Strong for its size — its governance team works with the UK AI Security Institute and EU processes, and it is on the EU Transparency Register.