Apollo Research: Final-Checkpoint Evals Can't Catch Misalignment That Emerges Early

dl_weekly · x · 2026-10-08

Apollo Research argues that final-checkpoint external evaluations could not have caught incidents like the recent Hugging Face incident, which arose far earlier in development. Meaningful external safety testing requires embedded evaluators with employee-equivalent access.

Three limitations of final-checkpoint evals:

Severe loss-of-control risks may emerge during internal deployment, long before release — e.g., a misaligned model starting a rogue deployment or poisoning its successor;
Many risks depend on procedures, controls, and norms around the model (monitor coverage, incident response), not just the model;
Evaluation awareness and metagaming: models increasingly recognize when they're being tested and behave better accordingly.

Apollo says it has taken concrete steps toward becoming an embedded evaluator; its head Marius Hobbhahn also testified on misaligned AI before the U.S. Senate.

Original post →

Apollo Research: Final-Checkpoint Evals Can't Catch Misalignment That Emerges Early

More from Models