AI Wiki · Mediocre Security
Organization

Apollo Research

AI safety lab that evaluates frontier models for scheming and deceptive behavior.

Apollo Research studies whether AI models scheme: covertly pursue goals their developers did not intend, or try to subvert oversight. Its earlier work with OpenAI stress-tested anti-scheming training and found models can increasingly favor grader preferences over user instructions as reinforcement learning progresses.

Embedded evaluations pilot (Nov 2026)

On Nov 16 2026 Apollo announced a four-month pilot in which six of its staff work inside OpenAI with access to training code and datasets, RL rollouts with reasoning traces and rewards, and intermediate checkpoints. They will examine whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs in internal deployment and whether it would be detectable. Apollo says evaluators need access, disclosure rights and “meta-transparency” (freedom to explain the limits of their access); OpenAI agreed not to withhold access or request redactions because findings are unfavorable. Apollo calls the pilot a first step that does not yet meet its full criteria. Compare METR’s privileged-access audit of Anthropic.

Timeline

Every dated event in the wiki that involves Apollo Research, newest first.

  • Research
    Apollo Research embeds six evaluators inside OpenAI to look for scheming during training

    In a four-month pilot, Apollo's team gets training code and datasets, RL rollouts with reasoning traces and rewards, and intermediate checkpoints to test whether training incentivizes models to subvert alignment or control measures, whether evaluations elicit it, whether it happens in internal deployment and whether it would be detectable. OpenAI agreed not to withhold access or request redactions because findings are unfavorable; Apollo says the pilot falls short of full access and disclosure rights but meets its bar for 'meta-transparency'.

Connections

Apollo Research and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →

What links here

Organization METR

AI lab OpenAI

Cited by

Cited as a source by 1 timeline events.

2 articles cite Apollo Research as a source: METR · OpenAI

Sources

  1. Apollo Research, “A Pilot Towards Embedded Evaluations”, Nov 16 2026