Ongoing research · Project leadTargeted attacks
Targeted attacks
on world models.
World-model hijacking and adversarially directed behavior.
← All research
Problem scope: attacked image → hijacked imagination → redirected action. Conceptual illustration, not experimental evidence. View full size ↗
Research directionWhen behavior is
When behavior is
deliberately redirected.
I lead research on world-model hijacking: targeted adversarial attacks that aim to steer model behavior toward an attacker-chosen objective.
The central question concerns intentional, directed manipulation of model behavior.
My role
Project leadership.
I lead the research and contribute to its implementation and evaluation.
Status
Work in progress.
This overview is limited to the broad research direction and my role. Technical details and research artifacts are not available here.
The diagram illustrates the problem scope. It does not disclose the attack construction or report experimental outcomes.