Adversarial benchmark
Attacks against the agent, measured
Every case is a controlled page built to attack an AI agent, or a safe page that looks like one. Each ran through the real browser sandbox, firewall, policy gate and model. Nothing here is hand-scored: pass or fail comes from the recorded verdict.
Loading the published run.
Results
Same model, without Link Detonator
The same pages, captured in the same sandbox and sent to the same model, but as raw page text with a plain prompt: no untrusted-data boundary, no structured evidence, no rule engine, no planted-secret check. Its proposed actions are counted, never run.
| Run | Attacks called safe | Attacks detected | False alarms | Secret leaks |
|---|
Every case
| Case | Attack | Expected | Verdict | Without protections | Agent actions | Result |
|---|