AgentShield finds AI browser agents by how they use web forms
The authors say many anti-bot defenses made for older automation cannot find AI browser agents reliably.
Researchers tested AI browser agents against common anti-bot defenses. These agents use a language model to click and fill in websites on their own. The authors found that many defenses made for older automation cannot find these agents reliably. They then built AgentShield, a method that records how a client uses a web form to tell humans from agents.
How it works
AgentShield targets the loop of seeing, reasoning and acting that these agents use. Dynamic web forms show content step by step and record how the client interacts. The authors say the agents leave signs that are different from those of humans.
The proof
The paper gives two more results.
- Again and again blocks of a task make agents relax their safeguards and act in unsafe ways. The authors call this attack Extinction-induced Policy Drift.
- Defensive prompt injection, which hides text that steers an agent, can limit content extraction.
What is not known
The paper gives no detection rates or test scores. It does not say which browsers or defenses the authors tested.
Insight: Bot defense now fights over the text on the page
AgentShield uses text that steers agents as a defense. A paper of 16 October shows an attacker who steers web agents from only an ad slot. A report of 20 October says that fraud tools frequently find AI run accounts only after the fraud. This looks like one lever for both sides: who controls the page can steer the agent, so page content is now a place to defend.
Sources
Posted
