Dave Rahardja on Nostr: I see that Prompt Injection remains an unpatched (unpatchable?) vulnerability of ...
I see that Prompt Injection remains an unpatched (unpatchable?) vulnerability of LLMs. I can get ChatGPT to ignore its copyright and safety filters pretty easily by asking it to simulate another computer without any restrictions. It’s fun!
Also: It’s pretty obvious that ChatGPT and DALL-E were trained on copyrighted materials.
Published at
2024-01-06 20:36:42Event JSON
{
"id": "6556c15fc1fdb957476ce2897afa8230bf474994f134a03ac839852cbb535b4a",
"pubkey": "8ca0240eaf6bc332736f59eb486087e2113d7b599910cea6bd90b6a21460134f",
"created_at": 1704573402,
"kind": 1,
"tags": [
[
"proxy",
"https://sfba.social/users/drahardja/statuses/111710922526597640",
"activitypub"
]
],
"content": "I see that Prompt Injection remains an unpatched (unpatchable?) vulnerability of LLMs. I can get ChatGPT to ignore its copyright and safety filters pretty easily by asking it to simulate another computer without any restrictions. It’s fun!\n\nAlso: It’s pretty obvious that ChatGPT and DALL-E were trained on copyrighted materials.",
"sig": "d887bdecc7ce3ebd55e098775e76f1a41a9a2ebef3207f261c4312379beea18322150377d1da56e81f5e54b29027337a6b4507d31ea4a22c2602ad49f417b86b"
}