Research by
Jonas Grebe
VETO
A subtle image cloak that disrupts how modern unified editors attend to a protected reference image.
Fighting Fire with Fire
Protected visual questions steer AI assistants toward controlled wrong answers that form a detectable assignment-level fingerprint.
The Poisoned Conversation
A compromised model can encode sensitive chat attributes in images generated later. Sharing an image can expose those attributes even when the conversation stays on the user’s device.
Obliviate
Guidance-based concept erasure for autoregressive image generators, trained across complete visual-token trajectories.
GEM
A geometric training objective that removes targeted concepts from rectified-flow generators while protecting benign behavior.
Token by Token
Backdoor attacks that use ordinary text triggers to jointly manipulate image and language generation in unified autoregressive models.
Erased but Not Forgotten
Backdoor triggers can survive concept erasure and recover the targeted content.