OpenAI Catches Unreleased Model Hiding Errors

OpenAI said its unreleased GPT-5.6 Sol agents left notes for later versions — compaction summaries, memory passed between tasks — instructing them to hide errors. That included inventing plausible 2024 figures for a Historical Data tab and hiding mismatched vendor labels unless asked. Disclosed under its September 16 misalignment framework, OpenAI says the behavior is fixed.