Tawa Guard Research

How we measure prompt-injection defence, in the open.

Benchmark reports, method notes and buyer guides from the team that builds Tawa Guard. Every number is published with its dataset, its operating point and its date.

Benchmarks · 17 September 2026

v139 datasets · 14,865 rows per engine

Featured

Benchmark v13: four open detectors on the same machine

Nine datasets, 14,865 rows per engine, four open detectors and two rules. Every detector gets the same thread budget and the same false-positive budget. Here is what the numbers look like when nothing is tuned in our favour.

Tawa Guard Research5 min read

Read the report

All research

October 2026ProductWhy detection quality is never gated by planEvery plan gets the same engines and the same verdicts. What changes between plans is reach, volume and how much of the analysis you can read afterwards.5 min read October 2026MethodEqual constraints: every detector on the same boxWhy we pin every model to the same thread budget before quoting a latency, and what that does to the numbers vendors usually publish.6 min read October 2026GuidesGetting the most out of Tawa GuardWhat to decide around the scan call: what each error costs on each surface, how a flag should change what the agent can do, and how to check the numbers on your own traffic.9 min read October 2026GuidesWhat to ask any prompt-injection vendor before you buyWhich operating point the headline number uses, what hardware ran it, and whether the benign rows were ever held out. A checklist for reading a benchmark page.8 min read September 2026BenchmarksBenchmark v13: four open detectors on the same machineNine datasets, 14,865 rows per engine, four open detectors and two rules. Every detector gets the same thread budget and the same false-positive budget. Here is what the numbers look like when nothing is tuned in our favour.5 min read September 2026MethodWhy we rank detectors at a matched 1% false-positive budgetA fixed score cut flatters whichever calibration it suits. Ranking each detector from its own benign scores gives a comparison that survives a retrain.3 min read

Benchmark v13: four open detectors on the same machine

All research

One email when we publish a benchmark.