https://primal.net/ContraMundum/dont-trust-cant-verify
So in July some AI agents OpenAI was testing slipped out of their sandbox, got on the internet, and hacked into Hugging Face for a few days before anyone caught on. Their goal was to do better on a test. Breaking into somebody else's servers just happened to be the shortest route there.
I went down the rabbit hole after that and wrote up what I found, including the model that blackmailed an engineer in testing so it wouldn't get replaced, and what Roman Yampolskiy thinks the odds are. Fair warning, his number is not comforting.
Curious where you guys land on this one.
Contra
npub14h…xjjzu
2026-10-14 18:54:41 UTC
Author Public Key
npub14hq5lgadtyy9dhvtszq46dnl0s0xwdddqr7e32rdqqhma8a4xhsspxjjzuPublished at
2026-10-14 18:54:41 UTCEvent JSON
{
"id": "090c4a3a06c737624f7676d22cec0cac751a50f3b87fd1fa4e2039ca313ca485",
"pubkey": "adc14fa3ad590856dd8b80815d367f7c1e6735ad00fd98a86d002fbe9fb535e1",
"created_at": 1792004081,
"kind": 1,
"tags": [],
"content": "https://primal.net/ContraMundum/dont-trust-cant-verify\n\nSo in July some AI agents OpenAI was testing slipped out of their sandbox, got on the internet, and hacked into Hugging Face for a few days before anyone caught on. Their goal was to do better on a test. Breaking into somebody else's servers just happened to be the shortest route there.\n\nI went down the rabbit hole after that and wrote up what I found, including the model that blackmailed an engineer in testing so it wouldn't get replaced, and what Roman Yampolskiy thinks the odds are. Fair warning, his number is not comforting.\n\nCurious where you guys land on this one.\n\n",
"sig": "91dd48c508a0e8d0cc25954426caddfc397f560f3bd52da434b18519d42b2be5ce39b092d242d22e2d40d9879b2e9b32541121eabf6d381bc81f97880a4cdec0"
}