AI

AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 06:18:25 UTC.

Briefings

AdvSim2Real co-evolves a task curriculum, an adaptive injection adversary, and a 4B web agent within a frozen web world model. The method rewards agents for solving half the tasks while penalizing them only for successful adversarial injections that flip judgments. Training in this simulator yields a 33.6% relative improvement in task completion against an unseen adversary compared to a base agent, with results verified on real browsers.

Key points

• The method co-evolves a task curriculum, an adaptive injection adversary, and a 4B web agent within a frozen web world model.
• Training in the simulator yields a 33.6% relative improvement in task completion against an unseen adversary compared to a base agent.

Context

Web agents complete user requests by reading and acting on pages written by third parties, making them vulnerable to injected instructions that redirect them. Current defenses fail against adaptive attackers, and existing adversarial training keeps tasks fixed after the agent solves them. AdvSim2Real addresses this by training agents and adversaries simultaneously in a simulated web environment.

Limitations

The study covers 150 web tasks and uses a frozen web world model simulator. Results are based on a 4B agent trained in the simulator and verified on a real browser; no additional background on the specific web platform or injection techniques is supplied.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 02:29:21 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeieeyx5ygmwtpfof6wovvf2ojcssmyiptqydpnyymyz65te6n4o3z4