新聞稿

Obiguard releases obi-lookout, a free open model that finds sensitive data in English and Malay text

A small open model, free under Apache 2.0, that finds personal and business-sensitive data in English, Malay and mixed text and runs on an ordinary CPU.

新聞稿以英文發布。

KUALA LUMPUR — Obiguard, the AI risk governance company, today announced the release of obi-lookout, a small open model that finds personal and business-sensitive data in text, so it can be masked or blocked before it reaches a language model or a log. It is free to use under the Apache 2.0 licence and is published on Hugging Face, with a quick-start guide on GitHub.

obi-lookout reads text and returns where each sensitive value is and what kind it is: names, ID numbers, phone numbers, email addresses, addresses, dates of birth, bank account numbers, money amounts, secrets such as API keys, internal hostnames, and client or project names. It was built for English, Malay and text that mixes the two, and handles Malaysian and Singapore formats. It is a detector, not a chat model: it returns positions and labels, not generated text.

The model is small enough to run on an ordinary CPU, with no GPU, and it works offline once downloaded, so the text it checks never has to leave an organisation’s own network. The quick-start repository shows how to use it from Python, from the command line, and as a Docker server that speaks the OpenAI API format, so tools and gateways that already call an OpenAI-style endpoint can call it too.

On a test set of 1,569 documents written from generated data, obi-lookout scored an overall F1 of 0.927, counting a hit only when the start, end and kind of a value were all exact. On documents containing nothing sensitive, 7.1% had at least one false alarm. Obiguard says plainly that the test set is synthetic and real documents will differ, that the model is not a guarantee that every sensitive value is found, and that it is weaker on organisation and client names and on text that is not prose, such as shell sessions and logs. The evaluation data is published so others can reproduce the numbers.

obi-lookout is built on the open-source GLiNER2 framework and fine-tuned from Fastino’s GLiNER2 privacy model. It was trained only on generated data, with no customer data and no real personal data. The model card lists its licences and attributions.

Security and engineering teams can use it to mask sensitive values before text is sent to an AI service or stored in logs, and to check what prompts and records contain. Because no detector is complete, it is meant to be one control among several. Obiguard also builds an AI governance platform that enforces policy on AI model calls and keeps an audit trail.

obi-lookout v0.1.0 is available now on Hugging Face. The quick-start guide takes a few minutes to run.

Update, 13 October 2026: the current release is obi-lookout v0.6.0. On real documents it now scores an overall F1 of about 0.90, up from 0.53 for v0.1.0. On the synthetic set above it scores 0.921, with 10.6% of documents containing nothing sensitive raising a false alarm. The figures in this announcement are for v0.1.0; the model card lists the results for each version.

關於 Obiguard

Obiguard 是一家於 2024 年在馬來西亞創立的深科技新創公司,獲得 Antler 支持。我們打造 AI 安全與治理基礎設施,協助組織檢視 AI 使用情況、落實政策,並保存監管機關所期待的稽核證據。了解更多:obiguard.ai/about。

媒體聯絡:[email protected]

← 全部新聞稿