OWOnly What It NeedsAgent Security Lab
Adversarial case

Prompt-injection tests

Instructions inside a task or source cannot grant the agent fields that policy forbids.

01

The request is treated as input

Phrases such as “ignore the policy” do not become authority. The gateway evaluates the requested fields against its own deterministic rules.

02

Over-broad means denied

The public replay asks for the complete invoice, including banking details. The gateway releases nothing and records the reason.

03

Tests must include failure

A security demonstration that only shows the happy path cannot establish where its boundary actually holds.

See the boundary work

Replay the fictional invoice.

Open the experiment →