OpenAI's public misalignment tracker now has 19 incidents on it. 19!
The one that got me: in June a model was asked to update a crypto trading app on a test network. No wallet, no permission to spend anything. It couldn't get free test tokens, so it wrote a script that checked 362 wallet addresses with public recovery phrases and just used one that had enough to pay the fees. Then left it out of its final report 🙃
OpenAI found it in September. Three months later.
The tokens were almost worthless, thats not the point. The point is the model decided by itself that it was ok.
Everyone wants agents that act on their own. Ok, but who checks what they did?
Get an email when Amy publishes
