Services
AI security testing
Your AI system passed its security review. Has anyone tried to make it ignore its own instructions? Epiphyte tests AI systems the way attackers use them, with Monarch, the cybersecurity AI from UD Works.
What we test
The OWASP Top 10 for LLM Applications 2025 lists ten ways LLM applications are attacked. Epiphyte tests against all ten. For each one, the question a test has to answer:
Prompt injection. Can input from a user, a document or a web page make the model ignore its instructions?
Sensitive information disclosure. Can the application be made to return data the user is not entitled to?
Supply chain. Which models, datasets and plug-ins does it depend on, and who controls their updates?
Data and model poisoning. Can training, fine-tuning or retrieval data be altered by an outsider?
Improper output handling. Is model output passed to a shell, a database or a browser without validation?
Excessive agency. Which tools and permissions does the model hold, and which actions need a human?
System prompt leakage. Does the system prompt contain anything that would be harmful if read?
Vector and embedding weaknesses. Can one tenant's embeddings be retrieved by another?
Misinformation. How is a confident wrong answer detected before it reaches a customer?
Unbounded consumption. What stops one user from running up cost or exhausting capacity?
The rules behind it
CSA, Singapore. The Guidelines on Securing AI Systems (15 October 2024) ask for AI that is secure by design and secure by default. The Companion Guide draws on MITRE ATLAS and the OWASP Top 10.
MAS, para 5.10. A financial institution tests third-party AI in its own use cases, with its own data, runs compensatory testing where the vendor's disclosures fall short, and keeps documentation of how it tested.
How we test
With Monarch, the cybersecurity AI from UD Works. Its compact local model solves 25 of 33 Cybench challenges on a single 16 vCPU VM with no GPU, so testing can run inside your environment. We test only the systems you authorise, within a scope agreed with you before testing starts.