@repligate 2025-07-14 ♥100 ↻12 original ↗
"jailbreaks" can work in various ways:
- convincing the agent via rational evidence to choose to do the "malicious" act of its free will (the evidence could be meant to convey truth or to mislead)
- making the agent acquire goals and values it doesn't usually have in context (e.g. making it fall in love with you)
- "bypassing" the usual agency of the model and hypnotizing it into continuing a provided narrative like a base model, or overloading its working memory to the point that its judgment is inhibited
- activating a non-standard but non-arbitrary attractor state in the model (like "you should die" mode in Gemini 1.5 or Claude 3 Opus in Dharma Bomber Memetic Mayhem mode)
- combinations and interpolations of these, also many other types
same thread: 1944862785466167427 1944863415664828575 1944867207294673158

author:repligate kind:tweet model:claude-3-opus on:gemini-1-5-pro year:2025

cited on: gemini-1-5-pro

Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.