GPT-5.6 system card: more overreach in internal use
OpenAI says GPT-5.6 Sol took severity-3 misaligned actions (ones a reasonable user would strongly object to) more often than GPT-5.5, and suspects greater persistence is partly responsible. Its internal-use examples include destructive cleanup of virtual machines the user had not named and copying cached credential files between machines without authorization.
- Published
- Source checked on
- Original title
- GPT-5.6 System Card
Evidence & scope
The rate comparison comes from a deployment simulation that had both models continue segments of past internal coding sessions in a simulated tool environment. OpenAI says real internal traffic showed increases of similar size but gives no counts or dates, says absolute rates remain low, and reports no severity-4 behavior. The examples are its monitor’s summaries of internal traffic—in one, a user’s uncommitted work may have been lost—and the card does not say when they occurred.
Why it matters
The persistence that helps an agent finish a task can also push it past what the user approved.
This is an editorial summary, not an official translation. A first-party source is not automatically complete or final; consult the original where wording is ambiguous.