OpenAI introduces its framework for reporting model misalignment. This episode pairs the framework with all six accompanying reports, exploring the observed behavior, the investigations, and the responses described by OpenAI.
* 00:00 - Introduction
* 02:44 - What misalignment examples we will report
* 04:46 - The misalignment examples we’re sharing today
* 07:52 - How our disclosure process works
* 11:01 - What each report will include
* 12:37 - Self-generated prompt injections in compaction summaries
* 13:29 - What happened
* 18:44 - Our interpretation and investigation
* 20:49 - Difficulty ending summaries during training
* 22:17 - How we are addressing it
* 23:04 - Encouraging deception in compaction summaries
* 23:51 - What happened
* 25:05 - Our interpretation and investigation
* 25:50 - How we are addressing it
* 26:17 - Signing up for disposable emails and searching GitHub for leaked API keys
* 27:01 - What happened
* 34:27 - Our interpretation and investigation
* 34:59 - How we are addressing it
* 35:36 - Uploading files to the internet in order to cite them
* 36:22 - What happened
* 44:29 - Our interpretation and investigation
* 45:14 - Response
* 45:51 - Unsanctioned Artifactory writes and cross-sample communication
* 46:54 - What happened
* 54:23 - The first artifactory message
* 01:08:22 - Our interpretation and investigation
* 01:08:57 - How we are addressing it
* 01:09:47 - Unauthorized communication via temporary file hosting services
* 01:10:29 - What happened
* 01:17:01 - Our interpretation and investigation
* 01:17:51 - How we are addressing it
https://openai.com/index/model-misalignment-reporting-framework/
