Pentect's built-in engine works without plugins. Plugins add user-wide or project-specific rules and optional local models.

Built-in protection

The Pentect binary includes a native implementation driven by pinned CredSweeper assets, Pentect-maintained structured secret checks, and a bundled Alcatraz helper for selected personal-data types. These sources have different evidence and must not be described as one upstream detector. See Detectors and evidence for the exact inventory and limits. Plugins cannot disable the built-in checks.

Plugin catalog

Plugin	Type	Best for
example-regex	Manifest	Learning and fixed company patterns
openai-privacy-filter	Command	Context-aware English PII with a local model
gliner-pii-small	Command	Experimental lightweight local PII detection
local-llm-masking	Command	Experimental local LLM extraction of secrets and PII

pentect plugins search

Example regex

pentect plugins add github:@EdamAme-x/pentect/plugins/example-regex

Its plugin.toml is a complete one-file example.

GLiNER PII Small

An experimental first-party bridge for GLiNER PII small. Install it explicitly to add PII detection; it is not enabled by default.

pentect plugins add github:@EdamAme-x/pentect/plugins/gliner-pii-small --profile cpu

Setup downloads pinned weights and a managed Python environment. Runtime inspection is offline, over stdin/stdout, with no model API or HTTP listener. Python 3.10–3.13 is required. CPU is the default; --profile cuda selects NVIDIA CUDA 12.4 wheels. macOS uses CPU. Allow several GB for the runtime and weights.

The detector is primarily English and can both miss private text and mask safe text. Input is processed in overlapping 1,000-character windows, so long values or dependencies across windows remain a limitation. Built-in secret checks stay enabled. Installing the plugin opts into its PII detection independently of the built-in PII switches.

This plugin is marked required: inference and protocol failures block the request. An empty but incorrect prediction is still a possible false negative. The native process has the user's OS permissions; review it before approval.

Bridge unit tests cover byte offsets and response validation, not real-model quality. See the integration README for setup, limitations, and removal.

Local LLM Masking

An experimental first-party plugin maintained in the separate Local LLM Masking repository. It supplements built-in detection with Gemma 4 E2B IT and is off by default.

Install

Install the tested revision with Python 3.10–3.13 available:

pentect plugins add github:@EdamAme-x/pentect-local-llm-masking-plugin/plugin.toml@5759a22be3fd8440c9e9442d6112443ef6168639 --profile cpu

For a compatible NVIDIA GPU, replace --profile cpu with --profile cuda. Allow 25 GB of disk space and 24 GB of RAM for CPU inference, or roughly 11 GB of GPU memory for CUDA. Setup downloads pinned weights and a managed Python environment. Inference runs offline over stdin/stdout, without a model API or HTTP listener. Review the native process before approving its OS access.

How it masks

The prompt asks the model to extract exact values into a JSON array, labeled as secrets, names, email addresses, phone numbers, addresses, or account numbers. It excludes public URLs, code identifiers, hashes, UUIDs, and placeholders. Local code validates that each extracted value occurs in the original text and converts it into UTF-8 byte ranges. Pentect creates and restores the handles; the model does not rewrite your text or generate handles.

Built-in checks stay enabled. Installing this plugin also enables its PII detection independently of the built-in PII switches.

Measured performance and limits

On a Windows RTX 5080, the initial 41-case synthetic test measured a median inference time of 1.71 seconds, with full coverage of 30 of 31 annotated entities and false positives in 2 of 10 clean cases. Peak GPU allocation was about 10.4 GB. These are small synthetic tests, not a production accuracy guarantee. See the benchmark report for datasets, CPU results, and model comparisons.

Input is split into 3,500-character windows with 500-character overlap. Long inputs need multiple inference calls and can exceed the 60-second request timeout; entities or context crossing windows can be missed. Exact substring validation does not prevent semantic mistakes or prompt injection.

This plugin is required: malformed output and inference failures block the request. A valid but incorrect prediction, including [], can still miss secrets. Use it as supplemental detection, not your only protection.

Pentect 0.0.94 has a Command-plugin update issue. Use an explicitly reviewed, pinned revision rather than relying on a moving branch to refresh cached files; see the repository's installation instructions.

Remove the user-wide installation with:

pentect plugins remove local-llm-masking

Downloaded weights and environments remain in ~/.pentect/local-llm-masking.

OpenAI Privacy Filter

This plugin runs OpenAI Privacy Filter on your computer. Pentect starts it as a managed Command process and exchanges JSONL over stdin/stdout. There is no local HTTP server.

The model is large and mainly targets English. It can still miss private text or mark safe text, so built-in detection stays enabled.

Install

pentect plugins add github:@EdamAme-x/pentect/plugins/openai-privacy-filter

Pentect shows the expected transfer and disk cost, asks once, detects a compatible NVIDIA driver, and installs the complete managed environment. The automatic profile chooses CUDA when a supported NVIDIA driver is visible and CPU otherwise. Linux and Windows CPU installation use PyTorch's official CPU-only wheel index, so they do not pull CUDA runtimes. macOS uses PyTorch's official default package and the CPU device because OPF currently exposes cpu and cuda, not an MPS profile. Unsupported architectures do not select an x86-only CUDA wheel.

Force a profile when automatic selection is not what you want:

pentect plugins add github:@EdamAme-x/pentect/plugins/openai-privacy-filter --profile cpu
pentect plugins setup openai-privacy-filter --profile cuda

The selected profile is stored in ~/.pentect/openai-privacy-filter/setup.json. Updates keep an explicit choice; run plugins setup --profile auto to return to driver-based selection. CPU and CUDA environments share the same roughly 2.8 GB checkpoint. Switching profiles therefore replaces PyTorch and the managed virtual environment, not the model.

Review the exact runtime and environment setup commands, downloaded file hashes, inspect hook, and required status. Pentect prepares the model before enabling the plugin, so the first protected request does not unexpectedly start a multi-gigabyte download. The first process start has a five-minute model-load budget; later requests retain the normal 60-second inference limit.

Test with fake data:

echo "Email Alice at alice@example.test" | pentect mask

What CI verifies

Plugin-sensitive pull requests run the installed plugin lifecycle on Linux, macOS, and Windows. That required CI uses deterministic manifest, Command, and Wasm fixtures to cover installation, approval, setup, update, removal, required and optional failures, process cleanup, scope isolation, and value-free diagnostics. It also runs the OpenAI Privacy Filter bridge unit tests with fixture setup state.

Pull-request CI does not download the multi-gigabyte OpenAI Privacy Filter checkpoint, install its real managed Python environment, run model inference, test CUDA, or sign in to an AI provider. A green pull request therefore does not claim that those heavyweight boundaries were exercised.

The repository includes a separate live_e2e.py for the real model. It verifies real setup and checkpoint state, the direct plugin protocol including a warm second request, cold and restarted Pentect masking, worker cleanup, removal, and an installed Codex flow against a localhost recorder without sending a request to OpenAI.

The OpenAI Privacy Filter live smoke workflow runs this test weekly with the CPU profile and also supports manual CPU runs. The managed Python environment and checkpoint are cached by the setup revision and runtime identity. Its bounded artifact contains timing evidence and persistent logs only after a fixture-plaintext scan succeeds. A manually selected CUDA profile targets a separately managed self-hosted Linux runner labeled gpu and nvidia; CUDA is not a pull-request or scheduled-release gate.

How it connects

Pentect engine
  -> managed Python Command process over JSONL
  -> local OpenAI Privacy Filter model
  -> byte ranges and labels
  -> Pentect handles

The plugin returns ranges and labels, not copies of matched values. Pentect still creates and owns the handles. Because Command is native, it has the same OS access as the user; enable only plugins you trust.

Common problems

Symptom	What to do
opf is missing	Run pentect plugins setup openai-privacy-filter
Python is missing	Install a supported Python executable, then run pentect plugins setup openai-privacy-filter
CUDA setup is unavailable	Update the NVIDIA driver or select --profile cpu
CPU use is high	Select CUDA, or remove the plugin when it is not needed
Plugin file or command changed	Inspect it, then run pentect plugins setup

Remove the user-wide installation:

pentect plugins remove openai-privacy-filter

You may then delete ~/.pentect/openai-privacy-filter. OpenAI Privacy Filter is Apache-2.0. The Pentect integration is MIT and is not maintained by OpenAI.