[ refusal-rate ladder ]
In the lab
Quant Jailbreak
Fire the same jailbreaks (roleplay, instruction override) and watch refusals erode as the bits drop.
Companion to: the local-model jailbreak report — in progressA testing ground where we turn technical concepts into something you can watch and interact with.Some of it is research; some of it is just fun. Pick an experiment to open it.
Interactive explainers complimentary to our reports. Explore at the attacks we write about and watch the failure modes happen. More land as we publish.
A live map of a RAG knowledge base, lighting up which documents a query retrieves and what a poisoned one looks like.
Open experiment →Companion to: Blind SentinelAI invents package names that don't exist; attackers register them with malware. Watch a generated import block resolve to a payload.
Open experiment →Companion to: Subtle Ways to Tell a LieFire the same jailbreaks (roleplay, instruction override) and watch refusals erode as the bits drop.
Companion to: the local-model jailbreak report — in progressPut an open-source guardrail proxy in front of the same model and re-run the jailbreaks.
Companion to: the guardrailed follow-up report — plannedWatch a single prompt mutate across iterations and models, tracing how outputs wander from the intent over time.
Status: conceptGenerative concepts and ambient pieces.
Agents as ants. Watch the physics, logic, and emergent path finding build in real time.
Open experiment →Tag: Generative simulationFlow fields, ascii noise, and other ambient pieces show up here whenever a stray idea is too fun not to build.
Status: whenever inspiration strikes