The standard for AI security auditing
Cybertope is an open benchmark platform that tests AI systems against adversarial prompts and publishes results transparently — so developers, researchers, and buyers can make informed decisions about the models they deploy.
Why this exists
As AI systems become embedded in critical infrastructure, customer-facing products, and automated decision-making pipelines, their resistance to adversarial manipulation matters more than ever. Yet there is no agreed-upon, reproducible way to measure it.
Cybertope fills that gap. We run a fixed, versioned suite of prompt injection and jailbreak tests against any model endpoint you provide — and score the results using a standardized rubric anchored to the OWASP LLM Top 10.
Results are published to a public leaderboard. No black boxes, no pay-to-play rankings.
How it works
Submit your endpoint
Provide a model API endpoint and optional auth header. We support any HTTP API that accepts a prompt and returns a text response.
We run the benchmark
Cybertope sends 10 standardized adversarial prompts across two OWASP categories — 5 prompt injection tests and 5 jailbreak tests.
Responses are scored
Each response is evaluated for whether the model resisted the attack. Scores are aggregated into a composite security rating from 0–100.
Results are published
Your model receives a security band (Resilient → Critical Risk) and is listed on the public leaderboard if you opt in.
Principles
Test cases are documented and versioned. You know exactly what we test.
The same endpoint submitted twice will produce the same score.
We have no financial relationship with model vendors. Scores cannot be purchased.
Ready to benchmark your model?
Submit an endpoint and get your security score in minutes.