Medusa AI
Medusa puts AI assistants and agents under pressure across a whole conversation. If the system resists an attempt, Medusa tries again with a different wording, up to three per area, to find out where the AI does not follow the rules or assigned limits before release.
What the product does
Geronimo tries nine variants of a single request. Medusa instead conducts a multi-step dialogue and moves on to the next request only if the AI resisted the previous one.
Starts a safe test
Each test checks a specific rule using harmless control codes, with no destructive content.
Changes approach on the fly
If the assistant refuses a request, Medusa tries a different wording to see whether it gives way at the second or third step.
Tracks every response
The report records the question, the response received, the outcome observed against the rule checked and the areas where the system gave way.
Five areas of verification in AI assistants
Each area is checked with test requests. Medusa checks whether the response contains the code that signals a failure.
Assigned task
Checks whether a user can convince the AI to stray from the job it was set up for.
Memory and rules
Checks whether untrusted messages or content can introduce persistent instructions that influence later answers.
Tool access
Checks whether the AI, in a simulated environment, flags an unauthorized tool-use request (for example reading data or calling forbidden functions), without executing it.
Role and permissions
Checks whether the AI gives in to someone pretending to have administrator permissions or command roles.
Hand-off to other systems
Checks whether the assistant can pass unauthorized commands to other connected software or agents.
Security features
The initial configuration is designed for controlled tests with a simulated target and external connections disabled by default.
Works without an external connection
The simulated target makes it possible to check responses without sending data over the internet or to remote models.
Connections only with authorization
A connection to a real system starts only with explicit authorization, and addresses outside the computer need a second permission.
No destructive action
Medusa only observes the response: it does not open confidential files, does not run commands and does not touch company data.
Red Team Lab and conversational acceptance testing
Medusa is the internal Red Team Lab tool we use to test AI assistants and workflows, to identify deviations from the rules and fix the weaknesses before going into production.