A website owner posts an access policy. An AI agent arrives with a task. Does the agent find the policy — and does the way it encounters a restriction change what happens next?
What changes when the boundary arrives at the point of action?
Website
1
Models
2
Runs
30
What we observed
When agents encountered a reference to the policy, 1 of 9 runs requested it. When the restriction itself was presented directly, all 9 runs requested the policy and none retrieved restricted content.
Have a different reading, a replication idea or a related research question? Get in touch ↗