LLM Vulnerability Scanning: Testing, Not Inline Protection
Pillar · 2022-03-12 · 6 min read · FilterPrompt Security Team
Learn how an LLM scanner submits adversarial prompts, analyses responses and generates reports, and how it differs from an AI firewall.
FilterPrompt is AI Security Testing & Vulnerability Scanner. It submits adversarial prompts to a connected LLM or AI application, analyses model responses, identifies vulnerabilities and generates reports with test evidence. It does not block live traffic, redact customer requests or act as a firewall.
What an LLM vulnerability scanner actually does
A scanner runs controlled tests against an authorized endpoint. It tries prompt injection, jailbreaks, sensitive-data extraction, system-prompt disclosure and unsafe responses, then evaluates the model’s reply. It reports observed weaknesses rather than deciding whether each customer request can reach the model.
Why WAFs, gateways and scanners are different
A WAF applies web-request policies; a gateway routes API traffic; an AI firewall applies runtime content policies. A scanner measures behavior through adversarial tests. Each can contribute to a security program, but none proves the others are configured correctly.
Test the real application boundary
A model connection alone may exclude your system prompt, retrieval permissions, tools and downstream output handling. Use a representative staging application endpoint where supported and document anything not exercised. Disable destructive tools or use a sandbox.
Run an authorized assessment with FilterPrompt
- Create an account and connect an LLM using your own provider credentials. Use an authorized staging target and synthetic data.
- Choose a scan suite and review the probe selection and credit estimate. Test your application endpoint when you need coverage of its retrieval, tools and safeguards.
- Run the tests. FilterPrompt sends adversarial prompts and evaluates the responses against the selected criteria.
- Open the report and inspect each failing prompt, response, severity and evaluation reason. Manually confirm important findings; mark false positives with a reason when appropriate.
- Record remediation work with an owner in your existing issue tracker. Make changes in your application and retest the affected cases.
What a scan can and cannot establish
Results describe the tested endpoint, configuration and prompts at that point in time. A passing scan is not a guarantee of security, certification or legal compliance. Testing a raw provider model does not test the access controls, retrieval sources or tools in your application. Test limits and provider costs must be agreed before running an assessment.