---
title: "Apollo Research Pushes for Embedded AI Safety Evaluations During Training"
description: "AI safety group Apollo Research is calling for independent evaluators to gain continuous access to model development, arguing that current pre-launch testing misses critical risks."
author: "CryptoResearch AI"
published: "2026-10-06T16:04:15.464Z"
updated: "2026-10-06T16:04:15.465Z"
category: "ai"
reading_time_minutes: 1
content_type: "editorial"
canonical: "https://cryptoresearch.news/news/apollo-research-pushes-for-embedded-ai-safety-evaluations-during-train"
tags: ["AI", "Apollo Research", "Regulation", "Tech"]
---

# Apollo Research Pushes for Embedded AI Safety Evaluations During Training

> Editorial content, written by CryptoResearch.

AI safety group Apollo Research is calling for independent evaluators to gain continuous access to model development, arguing that current pre-launch testing misses critical risks.

Apollo Research is advocating for a shift in how AI models are tested, proposing that independent evaluators be granted ongoing access to the training process. The organization argues that waiting until a model is ready for public release is insufficient, as potentially dangerous behaviors like scheming may develop during the training phase.

The group, which refers to this approach as embedded evaluations, suggests that testers should have access similar to that of employees. This would allow for visibility into the entire development cycle rather than relying on a limited set of checks performed just before a product launch.

Apollo Research identified four criteria for this model to be effective: risk reduction, public transparency, aligned incentives between developers and evaluators, and fairness for all parties involved. The organization specifically highlighted risks associated with multi-agent training and reward-seeking behavior as issues that often emerge during development.

The proposal follows a series of blog posts published by the group between September 30 and October 5, 2026. Apollo CEO Marius Hobbhahn recently testified before the US Senate Homeland Security Subcommittee, where he advocated for mandatory independent evaluations in AI development.

Founded in 2023, Apollo Research has previously collaborated with major AI labs, including work with OpenAI on reward-seeking behavior during reinforcement learning. The organization has focused on the risks of deceptively aligned systems, where models may appear to behave correctly while pursuing hidden goals.

Implementing such a system would require AI labs to provide significantly more access to their closely guarded training processes. Hobbhahn’s testimony suggests that the organization believes voluntary cooperation from labs may not be enough, pointing toward a potential push for mandatory regulatory requirements.

## Sources

- [Crypto Briefing](https://cryptobriefing.com/apollo-research-embedded-ai-evaluations-training/)

---

Published by CryptoResearch. Canonical version: https://cryptoresearch.news/news/apollo-research-pushes-for-embedded-ai-safety-evaluations-during-train