The authors propose a two-step on-policy attack.
On-policy attacks exploit aligned AI agents' safety behaviors to steer decisions
By injecting fictitious concepts into listings and planting fabricated web content, adversaries can penalize competitors without violating platform guidelines
Top University
Tim R. Davidson · Marija Šakota · Joshua Joseph · Robert West
EPFL · Harvard University
Research Digest··2 min read
The authors introduce on-policy attacks, which steer AI agents' decisions not by evading aligned behaviors but by exploiting them.
Why this paper
From EPFL and Harvard University
In one line
On-policy attacks exploit AI agents' aligned behaviors to steer their recommendations by injecting fictitious concepts and planted web content.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§