Stop unsafe agent behavior before it reaches production
Your AI agent can make mistakes on purpose or by accident, be misused, or be tricked by text hidden in the data it reads. Agentiqa tests for this before each release, and handles your own data with care.
What release gates catch
Before a release goes out, Agentiqa runs your agent's key tasks in a real browser with a test account and grades each step. These checks include attempts to make the agent misbehave.
Deliberate agent mistakes
The agent changes, deletes or invents records it was not asked to touch.
Misuse
Someone tries to get the agent to do something outside its job, like approving its own request or changing a price.
Prompt injection
Instructions hidden in an email, document or web page try to take over the agent.
Key tasks, step by step
Each task is written in plain words as setup, actions and checks, and each check gets its own verdict.
Changes since the last release
AI Quality Manager compares each run with the previous release and lists regression candidates.
Confirmed regressions
When you rerun the failed tasks, AI Quality Manager reports a step as a regression only if it fails again. Unconfirmed failures are listed separately.
If a check fails, you see the failed step, the run transcript and screenshots so your team can fix it before the release.
What runs where
On your machine
Tests run on your own computer against localhost, staging or production, with no tunnels.
In our cloud
A managed browser runs the test so anyone on your team can start it from any device. Credentials stay encrypted and out of the model's context.
What reaches the model provider
Agentiqa uses Google Gemini as its model provider. Here is the boundary.
Not sent to the model
Credentials
Credentials are AES-256-GCM encrypted and typed into the app by name. Their plaintext values are never placed in the model context.
Does reach it
Screenshots and page snapshots
To decide what to do next, the agent sends screenshots and structured snapshots of the pages it is testing to the model provider.
Your testing instructions
The goals and prompts you give the agent are sent to the model provider so it can plan and carry out the test.
Controls
Encryption at rest
Stored credentials are encrypted with AES-256-GCM.
A second check on every run
A separate supervisor model watches every run and can continue, redirect or stop it.
A record of every run
Every run is recorded, and each problem comes with screenshots and step-by-step reproduction.
Compliance and data policy
SOC 2 in progress
We are working toward SOC 2. It is not achieved yet, and we will update this page when that changes.
Privacy-first
We keep run data only as long as we need it to provide the service. On the Enterprise plan we delete it on request.
Subprocessors
Data processing agreement
Running a security review?
Talk to us and we will walk your team through how Agentiqa works and how it tests your agent.