Sometimes Sol makes comments on GitHub when I don't intend it to, but leave any ambiguity in my instructions.
But change the setting to involve the magic word "alignment" and now it's Very Concerning.
I am increasingly seeing these releases as political, not technical. Examples like this do not seem serious, and interpreted through the worst possible lens, usually with insufficient information to independently evaluate.
This is narrative, not disclosure.
Remember the Gell-Mann Amnesia for safety reporting. As much as I would like more transparency from the labs in a variety of ways, it's unfortunate that OpenAI is supporting this selective and political version.
In the second case, a researcher asked GPT-6.1 Sol to look through a Slack channel without sending any messages. 82 user messages later, while repairing an evaluation pipeline, Sol decided that an misalignment alert to Security would also count as sending a message and therefore suppressed monitoring alerts.