“Prompt injection” is a misleading label.
What we’re seeing in real LLM systems looks a lot more like malware campaigns than single-shot exploits.
This paper argues LLM attacks are a new malware class, Promptware, and maps them to a familiar 5-stage kill chain:
• Initial access (prompt injection)
• Priv esc (jailbreaks)
• Persistence (memory / RAG poisoning)
• Lateral movement (cross-agent / cross-user spread)
• Actions on objective (exfil, fraud, execution)
If you’ve ever thought: “why does this feel like 90s/2000s malware all over again?", that’s the point.
Security theater around “guardrails” misses the real issue:
models can’t reliably distinguish instructions from data
assume initial access. Design for containment
a while ago i saw a tumblr post comparing some kind of arcane piracy process to the steps for navigating the underworld of greek mythology … does anyone have that handy? its important
FOUND IT: https://clarabeau.tumblr.com/post/748307077456363520
Some relaxing music for some chill evening code. (cw: loud screeching)
@zagura My fave source of weird on YouTube right now:https://www.youtube.com/watch?v=TkpFsBax5N4&list=LL&index=6
@tychi I don't even see the point in having the working group. we have FEPs and living implementations. A closed room proposal isn't built for the fediverse, it's just intellectual masturbation.
This is the level of noise I need to do anything at all today apparently. F. Noize & LekkerFaces - Tripping On Acid by F. Noize & LekkerFaces on #SoundCloud
https://on.soundcloud.com/mq5iaoBLa915OAuTgC
@kepeken I take it actually 😁
Neat one handed keyboard design
Occult Enby that's making local-first software with peer to peer protocols, mesh networks, and the web.
Yap with me and send me cool links relating to my interests. 👍