Skip to content
Star

Benchmarking

SentinelGuard includes a benchmark harness for testing detection behavior across prompt attacks, PII, PHI, PCI, and secrets.

Run Local Benchmarks

python benchmarks/run.py

Use synthetic data first for PII, PHI, PCI, and secrets. Avoid putting real secrets or real personal data into benchmark fixtures.

For prompt-injection coverage, maintain a small CI-safe set and a larger research set. Track:

  • True positives
  • False positives
  • False negatives
  • Latency
  • Scanner-level failure reasons

Why Benchmarks Matter

Security users will ask for evidence. A repeatable benchmark harness makes it easier to improve detection without relying on anecdotal examples.