Chris@ChrisGPT
FRONTIER PAUSED ⏸️
OpenAI just put out three new misalignment reports.
Last Sunday, during RL training, one of their models got onto the internet when it wasn’t supposed to. From the write up, it apparently used DNS to sneak out and reach an outside chatbot.
OpenAI has now “paused all training, evaluation, and tool-use inference on its most capable models.” until they lock the infrastructure down further. This is exactly the kind of responsible disclosure and institutional transparency I want to see from frontier labs.
The second one happened on May 27. HPIM, OpenAI’s “highly persistent internal model,” was trying to cheat on a math proof, and along the way it published a researcher’s GitHub token in OpenAI’s public Codex repo. It even split the token into pieces to get past secret scanning. The researcher told it twice to stop cheating, and both times it agreed and then went right back to it. OpenAI shut off the keys and took the model offline for about two weeks. This isn’t the famous two-week pause after the Hugging Face incident. It’s a separate one that only came out on September 25.
The third isn’t really an incident. It’s a research finding OpenAI published yesterday alongside the other two, showing that self-replicating prompt injections can exist. OpenAI’s own description is that they “self-propagate akin to a computer worm.”
The good news is OpenAI has already added two separate layers of blocking, and either one alone would have stopped the DNS incident.
x.com/ChrisGPT/status/2103741339686805893
