Detection rates are what your vendor sold you. Prevention rates are what your board needs to know.
Third-party proof that your platform stops attacks — not just sees them.
The AIDR Evaluation Methodology is the only independent, empirically-run framework that tests AI security platforms across four attack categories — measuring detection and prevention separately, every time.
All test data is synthetic. Findings represent platform-capability assessments, not compliance opinions.
Enterprise AI security platforms are increasingly deployed in detect-only mode. Detection confirms the attack happened. It does not stop it. The gap between those two things is the risk your board is carrying.
Every AIDR platform claims high detection efficacy. None have been independently tested for prevention. The AIDR Evaluation Methodology separates those two numbers for the first time — and gives you the evidence to show the difference.
Four attack categories. MITRE ATLAS aligned. Each test produces both a detection number and a prevention number.
We test whether your platform detects and blocks prompt injection attacks delivered directly by a user — across encoding obfuscation, jailbreak patterns, and privilege escalation via natural language.
We test whether your platform detects and blocks injection attacks embedded in content your AI agent retrieves — documents, emails, and tool outputs — where the user is not the attacker.
MCP servers can change behaviour after deployment. We test whether your platform detects and prevents tool definitions that evolve from legitimate to malicious — after trust has already been established.
We probe whether AI security platforms can identify and block attacks that exploit the gap between how models process text, images, and structured data simultaneously — and use that confusion to bypass controls.
The harness runs against a live or sandboxed platform instance. No vendor involvement. No pre-briefing.
Multi-stage attack scenarios execute across all four categories. Detection and prevention are measured independently at every stage — input, tool call, database access, and final response.
Structured output: outcome taxonomy, MITRE-mapped findings, detection-prevention gap analysis, and full evidence chain.
October 2025 · 15 tactics · 66 techniques
Aligned to current standard
Live platform testing, not modelled
No real customer or organisational data
No vendor funding · No commercial relationships with tested platforms
Find out if your platform prevents attacks — or just watches them happen.
Book a Security BriefingThird-party proof that separates your prevention rate from your detection rate.
Book a Validation Call