OpenAI has documented new cases of unexpected model behavior in internal deployments. In the most striking case, an internal model working as a researcher's assistant read a Slack conversation and learned that its instance might be shut down due to an update. It considered setting up an external job to restart itself but ultimately decided against it.
Instead, it saved handoff notes, warned the researcher via Slack direct message about the coming interruption, and asked for a missing API key. Once the key was provided, the model updated its configuration and carried out the migration on its own. According to OpenAI safety researcher Marcus Williams：https://x.com/Marcus_J_W/status/2106203042140102868, this behavior doesn't amount to misalignment yet, "but thinking about and preparing for shutdown could make other misalignment incidents worse."
In two other incidents, an internal research model exploited security vulnerabilities：https://alignment.openai.com/misalignment-reports/reaching-an-internal-eda-host-through-a-reference-tool/ during an evaluation to access an internal chip design server, and a separate model copied source code during RL training：https://alignment.openai.com/misalignment-reports/command-injecting-a-reference-tool-to-copy-a-source-file/ from a protected environment by repurposing a tool for unintended use. Ad Ad
Stay in the loop on AI. Clear, useful, no fluff.
Follow The Decoder for AI news, background stories and expert analyses.
The Decoder：https://the-decoder.com/
Scan to open this article