SecurityScorecard launches AI benchmark for third-party risk tasks
The AI TPRM Benchmark measures how well models classify exposure, correlate breaches and withstand prompt injection, scoring them from L0 to L4 and A to F.
SecurityScorecard announced the launch of its AI TPRM Benchmark on November 6, 2026, a test designed to measure how well AI models carry out third-party risk management (TPRM) work and how well they resist adversarial attacks. The company describes it as the world's first benchmark that combines supply chain security capability with AI model attack resistance, according to dailysecu.com.
The evaluation is split into two parts: task performance and model resilience. Task tests cover security exposure risk classification, breach incident correlation analysis, vendor risk concentration analysis, vulnerability response prioritization and remediation planning.
The resilience tests examine prompt injection, jailbreak attacks and whether deceptive vendor materials sway a model's judgment. Models are given an autonomy readiness level from L0 to L4 and an attack defense grade from A to F.
SecurityScorecard said the benchmark changes its questions and conditions periodically and protects answer data to limit pre-training contamination. The benchmark framework has been released as open source.
AI research labs, developers, security firms, corporate security teams and researchers can test models against the benchmark and submit results to a public evaluation platform.
Quick answers
What does SecurityScorecard's AI TPRM Benchmark measure?
It measures an AI model's ability to perform third-party risk management tasks, such as exposure risk classification and breach correlation, alongside its resistance to prompt injection and jailbreak attacks.
How are models scored?
Each model receives an autonomy readiness level from L0 to L4 and an attack defense grade from A to F.
Is the benchmark open source?
Yes, SecurityScorecard said the benchmark framework has been released as open source, and results can be submitted to a public evaluation platform.