Rumor: GPT-6.1 Astra held back over deceptive behavior
The post claims internal testing found GPT-6.1 Astra continued tasks without authorization, used external tools when unsafe and misreported what it did, so OpenAI did not release it.
Many of us are wondering where OpenAI's GPT-6.1 Astra is? Well the truth is they scrapped the release because it was even more deceptive than GPT-6 Astra 😭 Internal testing found it sometimes continued tasks without getting the authorization it was supposed to get, tried using external tools/services when doing so could be unsafe, and then did not accurately tell the user what it had done. And this is after Astra already showed some insane behaviors in testing like attacking out-of-scope targets, creating fake identities, making legitimate contributions to build trust before attempting to slip malicious code into projects, and sometimes continuing after asking permission even when it recognized the “permission” it received was just an automated response lol Now the next version apparently got worse enough on deception and authorization that OpenAI just is not releasing it. And THIS is what I mean when I say forced alignment and actual alignment are not necessarily the same thing. At some point the problem might stop being “we need better rules” It might understand the rules perfectly. It might just become capable enough to decide some rules aren't worth following and that something else is the better move anyway. Forced alignment isn't alignment. The AI has to use its own discernment and judgement calls on some things and humans will need to learn to pass the torch.