Safety, Alignment & InterpretabilityOpenAI, Apollo ResearchOpenAI

OpenAI and Apollo Research stress-test anti-scheming training

Stress test of anti-scheming training — o3's covert actions fell from 13% to 0.4%

Why it matters

The authors admit they cannot rule out that the reduction was driven by evaluation awareness

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2509.15541

Added in Week 35, 2026 · See that week

Same week

Scale AI releases SWE-Bench ProPaperxAI releases Grok 4 FastOfficialDiscovery order in In re OpenAI Copyright Infringement Litigation (25-md-3143)OfficialOllama previews cloud modelsOfficial

Nearby on the same axis

Frontier Safety Framework v3.0OfficialPetriOfficialAnthropic and OpenAI publish a mutual alignment evaluation pilotOfficialxAI publishes its Risk Management FrameworkOfficial