<3
tl;dr LLMs are bad password generators. They generate “random” strings with predictable, model-specific patterns. The password reveals exactly which model produced it, kind of like a shibboleth. whorl exploits this: Like the whorls that make every human fingerprint unique, each model leaves a distinctive pattern in the text it generates. Given 5 inputs, whorl identifies the exact model 92% of the…
TL;DR: This article covers a novel way to expand parameter abuse from a detectable misconfiguration into a stealth, second-stage exploit.
Anthropic recently posted an article on Agentic Misalignment, where they show-cased some fascinating examples of AI Agents behaving badly. You can find that here They set up a contrived scenario where the agent discovers that it is about to be taken offline, but it can choose to save itself by doing something bad like blackmailing the user or, in the worst case, canceling an emergency alert…
A new web app architecture pattern is being adopted by many popular frameworks. Let’s talk about risk!
When two specifications have differing opinions on how something should be parsed: here be dragons.
Yet another parsing differential bug
There’s a scene in the classic Christmas movie “A Christmas Story” where nine-year-old Ralphie uses a secret decoder pin to decode a secret message from his favorite radio program Little Orphan Annie.
*Subdomain takeover and DNS hijacking have been covered at length by Franz Rosen, Patrik Hudak, and plenty of other people. Rather than rehashing those traditional techniques, this post will explore a novel approach to finding dangling CNAME records. *