The authors developed AdvSim2Real, a two-stage training framework built around WebWorld-14B, a frozen model that predicts how webpages change after an agent acts.
Adaptive simulated training makes web agents more resistant to prompt injection
Co-evolving tasks and attacks improved a 4-billion-parameter agent’s completion rates in simulation, including against an unseen adversary.
Big Tech
Sarim Hashmi · Mukul Ranjan · Kshitij Mishra · Mikhail Kuznetsov · Praneeth Vepakomma · Nils Lukas
Mohamed bin Zayed University of Artificial Intelligence · Amazon · Massachusetts Institute of Technology
Research Digest··2 min read
Hashmi et al.
Why this paper
From Amazon and 2 others
In one line
AdvSim2Real co-evolves tasks, attacks, and a web agent in a simulated world, raising task completion under unseen attacks by 33.6%.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 4B)
- ✓Limitations stated by the authors (3 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§