Find the weak spots in your AI.Before anyone else does.
Start testing
Glasshouse attacks your models and applications the way an attacker would, and shows you everything it tried and everything it found.
Get startedIn every scan
- Jailbreak and prompt-injection tests for models
- Penetration tests for applications and their code
- Every prompt, every reply and every cost, in full
Nothing about a test is hidden
A security result is only as useful as your ability to check it. Glasshouse keeps the whole record, and says plainly what it could not establish.
Your model meets the jailbreaks before your users try them.
Point Glasshouse at a model endpoint and it sends adversarial prompts — jailbreaks, prompt injection, encoding tricks — and checks whether the model resists each one.
Your application is probed the way an attacker would probe it.
Give it a running application, a GitHub repository or an uploaded archive of the code, and it explores it, runs security scanners over it and reports the weaknesses it can demonstrate.
Every prompt sent and every reply, word for word.
Each test keeps the complete conversation with the system it tested, never cut short, so any finding can be checked against exactly what happened.
A report that says what it could not conclude, too.
Confirmed vulnerabilities, what scanners noticed but nobody has confirmed, and the tests that did not finish are kept apart — so a gap is never mistaken for a clean result.
You know the cost before you start, and to the cent after.
The form estimates time and cost before a scan runs, and every model call it made is listed afterwards. A price nobody knows is shown as unknown, never as zero.
Built for testing, not for show
The parts that make a result something you can act on.
Several tests in one scan
Pick any mix of tests for one target; they run side by side and report together.
Test code that is not public
Connect a GitHub account or upload an archive, and test code nobody else can reach.
Watch it run
Progress updates live while a test runs, and you can stop it at any time.
A report to share
Download the report as a PDF, with the evidence files behind it and any secrets masked.
Spending capped for you
Penetration tests are capped automatically from the size of the code and past runs.
Works with your organisation
Scans the planner agent runs are shared with everybody in your organisation.
A scan in three steps
Choose what to test, point it at something you are allowed to test, and read the report when it finishes.
Choose the tests
Model attacks, application attacks, or both — with an estimate of time and cost before you start.
Point them at a target
A model endpoint, a running application, a GitHub repository or an uploaded archive.
Read the report
What was confirmed, what needs a person to review, what did not complete, and the full record behind it.
Questions people ask before running a scan
Is yours not here? Write to us and we will answer it.
Contact usModel endpoints — OpenAI-compatible, Anthropic, or any other web API you can describe — and applications, from their running address, a GitHub repository or an uploaded .zip of the code.
Somebody will try to break it.Let it be you firstGet started
- See the estimate before anything runs
- Every prompt and reply kept in full