Essay casts prompt injection as social engineering of models
Original title: AI Security Issues Are Social Engineering Issues
The author argues capability-based security, as in CHERI and Cloudflare Workers, limits confused-deputy abuse of model instructions.
An essay compares prompt injection to social engineering, arguing that models follow instructions hidden in content the way people can be tricked. It says capability-based security gives a program only the authority needed for a task, limiting confused-deputy abuse. The author points to existing work including CHERI, Fuchsia, Capsicum, Cap’n Proto, Cloudflare Workers, and Spritely’s OCapN, and hopes AI-assisted reengineering will help adoption.
Unconfirmed: The essay’s “Black October” flaw wave is the author’s framing, not a documented incident in the text.
- The essay says smuggled instructions can range from style changes to requests for known data.
- It names Unix file descriptors and WASI as existing capability bubbles.
- CHERI encodes capabilities in hardware; Capsicum ports the model to existing software.
- Spritely, funded via the Filecoin Foundation, built object capabilities and OCapN.
- Organizations
- Spritely InstituteCloudflareEFF
From the original article
5 min read at oblomovka.comWhen I was at EFF in the 2010s, I spent a lot of time fretting about the fundamental security of networked devices. Apart from the usual concerns, I struggled to imagine a transition to a redecentralized future with fragile personal computers and servers. Peter Eckersley , my co-worker and frequent conversation partner, shared my concern, but had another reason to be worried.
By oblomovka.com by dannyob · oblomovka.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.