Dominic 🇪🇺 🏳️🌈 🇺🇦 on Nostr: It’s interesting to reflect on how processing natural language has vastly increased ...
It’s interesting to reflect on how processing natural language has vastly increased the attack surface of #AI compared to traditional software. Most injection attacks exploit weaknesses to run instructions that are completely unrelated to the nominal purpose of what is submitted (e.g. buffer overflows). But redundancy in natural language means it is possible to make minor changes to a prompt and bypass safeguards on the #LLM.
https://www.404media.co/apparently-this-is-how-you-jailbreak-ai/Published at
2024-12-20 10:22:00Event JSON
{
"id": "08d1fcf8ef41b8d44653b7eb5685d4af42eef2043c433902e87f190450c38a2d",
"pubkey": "bec3457bfb07cac68db001160f204c0ca2a0ee46d398fa62213a02e60f98c17b",
"created_at": 1734690120,
"kind": 1,
"tags": [
[
"t",
"ai"
],
[
"t",
"llm"
],
[
"proxy",
"https://mastodon.social/users/riotnrrd/statuses/113684651733531127",
"activitypub"
]
],
"content": "It’s interesting to reflect on how processing natural language has vastly increased the attack surface of #AI compared to traditional software. Most injection attacks exploit weaknesses to run instructions that are completely unrelated to the nominal purpose of what is submitted (e.g. buffer overflows). But redundancy in natural language means it is possible to make minor changes to a prompt and bypass safeguards on the #LLM. https://www.404media.co/apparently-this-is-how-you-jailbreak-ai/",
"sig": "e617ad92053ee83da3566d9225881407ac40806d252d3e112c7fa5c2a182b6dff536fc2543df235dd96f65708f03b81a17c4810a6c35191ac3cba51240928538"
}
