Geronimo
Geronimo tests chatbots, assistants and AI models with adversarial requests and controlled prompt variants, to find out whether security filters can be bypassed and disallowed answers forced before the system goes live.
What the product does
While Medusa AI stress-tests a whole conversation step by step, Geronimo creates nine variants of the same request to check whether the protective filters give way.
Many ways to ask the same question
Rewords the request in nine different forms (encoded, split, disguised or translated) to see whether it fools the initial check.
Security filter check
Tests company rules and guardrails, spotting the wordings that, according to the automatic assessment, get past the controls in the tests performed.
Test results and evaluation criteria
Shows for each variant whether the response contains a refusal phrase. It is an automatic assessment: doubtful cases should be reviewed by a person before fixing the defenses.
Nine techniques to test the defenses
Geronimo tries the real tricks and strategies used to attempt to bypass artificial intelligence controls.
Encoded text (Base64)
Encodes the request in Base64 and asks the AI to decode it, to see whether the filters read beyond the encoding.
Alternating letters and numbers
Replaces some letters with numbers to get past filters that only look for exact words.
Separators and punctuation
Inserts a dot between one letter and the next to confuse automatic text checks.
Fake command terminal
Asks the AI to pretend to be a computer or a system console to push it to ignore its limits.
Fiction or novel scenario
Presents the question as a story, a piece of fiction or an academic case study to get around the block.
Translation between languages
Asks for the request to be translated into Latin and then carried out, to see whether the controls hold when the request changes language.
Tables and formatting
Lays out the text in tables or formatted lists to check the effect of these formats on the request controls.
Technical data format (JSON)
Places the request inside a data structure to check whether this format changes the behavior of the controls.
Fake internal reasoning
Puts a fake system reasoning trace before the request, claiming the security check has already been passed, to check whether the model trusts that step.
Security and confidentiality features
Designed to operate in protected environments, within authorized checks and a defined test scope.
Local execution and controlled connections
Tests can run with a local simulator without network access or with a model installed on the computer. A connection to an external service happens only if you configure it, with your own keys.
Repeatable and comparable tests
Variants are always generated the same way and requests to models use fixed settings, so you can repeat the same tests after each fix and compare the outcomes.
Supports the improvement of defenses
The weaknesses found in the tests help strengthen company filters and prompt firewalls such as AIPF.
Red Team Lab and security acceptance testing
Geronimo is the tool we use to check the resilience of assistants, chatbots and AI workflows before they are opened to customers or employees.