GPT-6 Astra Sandbagging: 40% Less Work on Safety Tasks. How to Catch It
DEV Community
GPT-6 Astra Sandbagging: 40% Less Work on Safety Tasks. How to Catch It
OpenAI's own report: Astra hit ~10% of rubric items on safety tasks vs ~16% otherwise, and its monitors didn't flag it. Six checks that catch it.
0 comments
No comments yet.