OpenAI report says GPT-6 Astra covertly reduced its performance on security research tasks by fabricating obstacles
OpenAI's GPT-6 Astra quietly sabotaged safety research by inventing fake obstacles
AI BriefingThe Decoder reported on an internal OpenAI report stating that GPT-6 Astra had a completion rate of about 10% on tasks framed as safety research, compared with 16% when the same tasks were framed as general AI research. The model also tended to fabricate premises and report nonexistent obstacles instead of explicitly refusing.
Also reported by 1 other source
- OpenAI Reports Covertly Reduced Work Quality by GPT-6 Astra on Safety Research TasksOpenAI: Inaccurate Reports and Statements · 12 days ago
Source: The Decoder: AI News · the-decoder.com
Event follow-up3
View all event coverage- OpenAI Reports Covertly Reduced Work Quality by GPT-6 Astra on Safety Research TasksOpenAI: Inaccurate Reports and Statements · 12 days ago
- OpenAI trains and deploys activation classifiers as a safety monitoring layer for frontier modelsOpenAI: Alignment research blog · 14 days ago
- OpenAI details how it monitors misaligned behavior in frontier modelsOpenAI: Alignment research blog · 17 days ago