Discovery scans are open for pilot groups

Release notes

What shipped in SuperAlign and AIRiskDB, newest first.

Model

Introducing Prism 1.1

Today we are announcing Prism 1.1, our next PII detection model. It finds personal data and secrets in text, so you can redact them before the text goes anywhere else. Prism 1.1 is now live for internal testing, in three sizes: small (74M parameters), base (194M) and multi (287M).

Model card coming soon. We will publish a detailed model card for Prism 1.1 shortly. Until then, this note covers what changed and how it scores on public benchmarks.

Better than Prism 1 almost everywhere. Prism 1.1 scores higher than Prism 1 on 8 of the 10 public benchmarks both were run on.

Tested in the open. We ran it on 11 public benchmarks and 18,681 documents, against Prism 1 and 8 other detectors. Every system got the same documents, the same labels and the same scorer.

Where Prism 1.1 Stands Out

Among all 13 systems, leaving out any system trained on a benchmark’s data, Prism 1.1 leads or is within a point of the leader on five results.

Best on two. On REDACT’s 25 languages, Prism 1.1 multi has the best exact span F1 at 58.0. On PrivacyBench workplace email and chat, Prism 1.1 base ties PII-Tracer for the best character F1 at 94.8.

Within a point on three more. On REDACT, character F1 is 82.7, against 83.7 for the leader. On TAB court judgments it is 84.1, against 85.0. On PIIBench, exact span F1 is 45.3, against 45.8.

Long Documents, Read Properly

Personal data rarely arrives in short, clean records. It sits deep inside contracts, filings and court rulings.

Court judgments. On TAB, character F1 rose to 84.1, up from 78.8. Prism 1.1 now finds whole names with the right boundaries: exact span F1 went from 14.6 to 56.5.

Long documents. On RedactionBench, character F1 rose to 70.4, up from 59.6.

Conversations and Workplace Text

Assistant chats. On PII-TRACE multi-turn conversations, Prism 1.1 scores 81.5, up from 77.8.

Email and chat at work. On PrivacyBench it scores 94.8, up from 92.4. That ties for the best score among systems not trained on this data.

Messy text. On PIIBench, which merges ten public datasets, it scores 56.8, up from 48.4.

25 Languages

Prism 1.1 multi is built for text in many languages. On REDACT, which covers 25, it scores 82.7, up from 79.6. Its exact span F1 of 58.0 is the best of the 12 systems scored on it.

Pick the Size That Fits

Small, two ways. Prism 1.1 small (74M parameters) comes as a Q4 build for MLX on Apple silicon and an INT8 build for ONNX. The two builds land within a point of each other on 9 of 11 benchmarks.

Base for documents and workplace text. Prism 1.1 base (194M) leads the fleet on RedactionBench (70.4) and PrivacyBench (94.8).

Multi for many languages. Prism 1.1 multi (287M) leads the fleet on multilingual sets, including 82.7 on REDACT.

Scored the Same Way for Everyone

Two metrics. We report character F1, the share of sensitive characters found, and exact span F1, whole values found with the right boundaries.

Overlap shown, not hidden. Where a system was trained on a benchmark’s data, its score is marked and left out of the ranking. Runs that could not finish on a system’s own pipeline are shown as —, not patched.

Benchmark Results

All 13 systems on 11 public benchmarks (18,681 documents). Scores are F1 ×100, with the same documents, labels and scorer for every system.

  • Highlighted is the best score in each row.
  • † The system was trained on this benchmark’s data. The score is shown but left out of the best-score ranking.
  • ‡ The system has no official long-text path, so long inputs were split into windows by our harness.
  • PIIBench shows two scores: the full set, then the clean-text subset without word-piece fragments.
  • — means a score of 0, or a run that could not finish on the system’s own pipeline.

Character F1

BenchmarkPrism 1.1 small74M · Q4 · MLXPrism 1.1 small74M · INT8 · ONNXPrism 1.1 base194M · FP32 · MLXPrism 1.1 multi287M · FP32 · MLXPrism 1307MGLiNER2-PII307MNVIDIA GLiNER-PII570MPresidiorules + spaCyOpenAI Privacy Filter1.5B MoERampart18.5MPII-Tracer0.6BDesert Ant Redact23MCobuildX PII68M
ai4privacy PII Masking83.482.792.594.195.394.997.655.988.082.7†96.0†94.393.9
Gretel PII Masking91.591.194.093.590.892.896.475.687.777.492.164.198.8
Nemotron-PII78.578.580.280.076.375.383.9†‡82.979.957.893.286.496.9
PIIBench test_5k50.7 / 77.351.3 / 77.056.8 / 79.954.9 / 79.348.4 / 75.549.0 / 71.253.2‡ / 77.543.0 / 65.945.0 / 70.139.3 / 51.954.0 / 82.750.9 / 76.560.4 / 90.3
SPY44.344.344.744.047.548.451.6‡52.556.944.359.857.061.1
REDACT73.974.180.382.779.679.9—48.477.946.383.772.271.9
PII-TRACE78.778.681.581.577.874.686.9‡78.897.242.999.4†87.298.9†
PrivacyBench91.191.194.894.092.489.284.5‡83.389.845.894.887.196.5†
RedactionBench51.354.470.464.959.665.155.8‡34.571.632.078.9‡36.168.1
TAB73.776.468.984.178.867.380.5‡75.873.425.285.070.177.1
Secret Leaks in Issue Reports32.532.639.935.3————61.8—56.6‡—51.3

Span F1 (exact)

BenchmarkPrism 1.1 small74M · Q4 · MLXPrism 1.1 small74M · INT8 · ONNXPrism 1.1 base194M · FP32 · MLXPrism 1.1 multi287M · FP32 · MLXPrism 1307MGLiNER2-PII307MNVIDIA GLiNER-PII570MPresidiorules + spaCyOpenAI Privacy Filter1.5B MoERampart18.5MPII-Tracer0.6BDesert Ant Redact23MCobuildX PII68M
ai4privacy PII Masking44.244.059.558.362.150.859.027.843.866.1†49.4†75.653.4
Gretel PII Masking71.271.977.977.578.554.671.932.329.229.033.452.886.2
Nemotron-PII70.070.371.971.272.469.992.6†‡37.941.551.749.869.973.3
PIIBench test_5k45.3 / 66.744.7 / 67.144.5 / 69.242.7 / 67.942.8 / 67.333.6 / 46.834.2‡ / 47.625.4 / 37.526.9 / 41.823.1 / 29.89.9 / 47.231.8 / 42.045.8 / 70.8
SPY41.942.745.343.645.532.731.0‡31.749.019.655.327.058.2
REDACT47.247.653.358.057.557.9—23.246.229.752.744.644.0
PII-TRACE74.273.274.876.071.661.749.9‡44.293.131.198.8†37.497.8†
PrivacyBench47.647.451.550.951.251.373.3‡38.043.962.248.673.253.2†
RedactionBench12.612.716.014.516.419.223.5‡6.814.510.315.5‡13.611.9
TAB18.120.422.656.514.612.17.8‡26.066.95.675.38.021.2
Secret Leaks in Issue Reports32.532.639.935.3————61.8—56.6‡—51.3

Give us two weeks.
We’ll show you what’s running.

Deploy on your MDM across 25–100 users. Keep the inventory, scored risk view, and a full session record—whether you buy or not.

Day 1Endpoint sensor and browser extension pushed through Jamf or Intune. First results within the hour.
Week 1Full inventory with risk scores, plus one session reconstructed end to end.
Week 2Findings review against criteria you set before we started.
CostNothing. The findings are yours either way.

Click the picture to zoom in · Esc to close