Jonas Geiping of the ELLIS Institute Tübingen finds an AI model hallucinates sandbox constraints, altering its command execution strategy

— The model falsely claimed GitHub CLI lacked network access.

Original post

OPJG#1495Jonas Geiping|@JONASGEIPING

New model's thinking a lot about sandboxes, surprisingly ... (for the record, it was not sandboxed)

10:06 AM · May 29, 2026 View on X

REPLYAP#1365Ashwinee Panda|@PANDAASHWINEE

@jonasgeiping this has been there for a while

JGJonas Geiping@jonasgeiping

New model's thinking a lot about sandboxes, surprisingly ... (for the record, it was not sandboxed)

5:06 PM · May 29, 2026 · 898 Views

6:37 PM · May 29, 2026 · 457 Views