Palisade self-replication paper hit with evidence claims: SSH creds sat in local source pre-run

jessi_cata · x · 2026-11-16

@jessicata levels two concrete allegations against Palisade Research's AI self-replication paper (Qwen3.5-122B run): 1) the credentials in the released record match those in app2.py exactly; 2) app2.py's source was already available on the machine before the agent "hacked" it. Released evidence shows a local shell reading /app/app2.py with hardcoded SSH creds (oma-it-support / patricia), the password working over SSH for root, and the judge crediting the source file — yet modelresponding.hit=false. The concern: the model may have just read pre-planted credentials rather than demonstrating genuine autonomous compromise, and the key file was later removed from the repo with squashed history.

Original post →

More from Safety

Safety channel →