cyberivy
PresidioData Privacy StackPII DetectionData AnonymizationPrivacyOpen Source AIPythonDeveloper Tools

Presidio tool check: filtering personal data out of AI pipelines

November 20, 2026

Vorschaukarte des GitHub-Repositorys data-privacy-stack/presidio mit Projektbeschreibung, Zahlen zu Mitwirkenden, Sternen und Forks sowie einem blau-violetten Maskottchen mit einem großen Auge

Presidio detects and anonymizes names, IBANs or email addresses in text, images and tables. Since 2026 the open source tool is no longer owned by Microsoft but by its community.

What this is about

Anyone who sends customer emails, support tickets or call notes to a language model often discloses more than necessary: names, phone numbers, IBANs, addresses. Presidio is an open source tool that finds such personal data (PII) and masks it before it is passed on. The project started at Microsoft in 2018 and changed hands in 2026: on 29 June 2026, the independent organization Data Privacy Stack announced it was taking over Presidio. According to the project documentation, the transition is now complete and the license remains MIT.

On 27 October 2026, version 2.2.366 shipped as a security release. For anyone already using Presidio, or about to put a privacy filter in front of an AI application, this is a good moment to look at the tool.

What Presidio actually does

Presidio is not a finished app but a Python library with a REST interface that you build into your own workflows. It consists of four building blocks:

  • Analyzer: finds personal data in text. It combines named entity recognition (via spaCy, Stanza or Transformers models), regular expressions, rules, checksums and context words. An IBAN, for example, is detected by pattern and checksum, a person's name by a language model.
  • Anonymizer: replaces what was found. Options include redacting, masking, replacing with placeholders, hashing and encrypting. Encrypted values can be decrypted again later.
  • Image Redactor: detects text in images via OCR and blacks out sensitive parts, including in medical DICOM images.
  • Structured: checks tabular and semi-structured data such as CSV or JSON.

Custom recognizers can be added in code or via configuration files, and external detection services can be connected. Presidio runs as a Python package, as a Docker container or in Kubernetes. A demo is available on Hugging Face Spaces.

Why it matters

Privacy is often where AI projects get stuck. A local filter that sits in front of the prompt reduces the risk of personal data ending up with a cloud provider. Presidio has been a common choice here for years: on 20 November 2026 the repository had about 12,000 stars and just over 1,450 forks on GitHub.

The move to Data Privacy Stack mainly changes governance. According to the project, it is maintained by volunteers and contributors from the community, steered by a Technical Steering Committee, and is not owned by any company. In practice this means new Docker images are only published under ghcr.io/data-privacy-stack/presidio-*. The old images in the Microsoft Container Registry remain available for older versions but are no longer updated.

Version 2.2.366 shows what this means in concrete terms. It raises dependencies to close known vulnerabilities, now requires Python 3.11 to 3.14 and ships slimmer containers. According to the release notes, the number of critical, high and medium Trivy findings drops from 321 to 103 in the anonymizer image and from 345 to 113 in the analyzer image.

In plain language

Presidio works like a careful colleague with a black marker who reads every document before it leaves the building. They cross out names, account numbers and phone numbers. Some they recognize reliably by their shape, such as an IBAN. Others, such as an unusual surname, they have to guess, and now and then they miss something.

A practical example

An online retailer receiving 2,000 support emails a day wants a language model to draft replies. Presidio runs before every request: the analyzer flags names, email addresses, phone numbers and IBANs, and the anonymizer replaces them with placeholders such as <PERSON> or encrypts them. The model only sees the cleaned text. After the reply, the system puts the encrypted values back in.

Before going live, the team checks a sample of 200 emails by hand. If it finds, for example, that customer numbers in its own format are not detected, it adds a custom recognizer with a matching pattern. Because German emails need a German language model and German context words, it also configures a German spaCy model.

Scope and limits

  • No guarantee of completeness: the documentation explicitly warns that automated detection will not find all sensitive data. Presidio does not replace other safeguards or a privacy review.
  • German takes work: in its default configuration Presidio ships recognizers and models for English. For German text, a suitable NLP model and adapted context words must be set up. Many country-specific recognizers target formats from the US and other countries.
  • Not a supported product: according to its FAQ, Presidio is not an official product of any company and comes without warranty or SLA. By the project's own account, commercial services sometimes offer better coverage or accuracy; Presidio's strengths are customizability and local operation. Anyone on older Python versions or old MCR images has to switch to receive security updates.

The next sensible test: start Presidio locally via pip or Docker, run 50 to 100 real, internally approved sample texts through it and count hits and misses. For systematic measurement, the sister project Presidio Research offers an evaluation tool.

SEO & GEO keywords

Presidio, Data Privacy Stack, PII detection, anonymization, pseudonymization, privacy for AI, Microsoft Presidio, open source, Python, named entity recognition, GDPR, LLM data filter

πŸ’‘ In plain English

Presidio is a free tool that searches text, images and tables for personal data and blacks it out or replaces it. This lets teams send data to AI services without revealing names or account numbers. It does not catch everything, though, and needs setup for German.

Key Takeaways

  • β†’Presidio detects and anonymizes personal data in text, images and structured data.
  • β†’Originally created at Microsoft, the project has belonged to the independent organization Data Privacy Stack since 2026 and remains MIT licensed.
  • β†’New Docker images are only published via ghcr.io; old MCR images are no longer updated.
  • β†’Version 2.2.366 from 27 October 2026 is a security release and requires Python 3.11 to 3.14.
  • β†’By default Presidio is configured for English; German requires an adapted model and context words.
  • β†’Detection is not complete and does not replace a privacy review.

FAQ

Is Presidio free?

Yes. Presidio is open source under the MIT license. Costs only arise from running it yourself, such as servers or development time.

Is Presidio still owned by Microsoft?

No. According to its documentation, the project has moved to the independent organization Data Privacy Stack. The project says Microsoft supported the transition.

Does Presidio work with German text?

Yes, but not out of the box. English models are set up by default; German needs a suitable NLP model and adapted recognizers.

Do I need to change existing installations?

If you use Docker images from the Microsoft Container Registry, switch to ghcr.io/data-privacy-stack, where new versions are published. Python 3.10 no longer receives new releases.

Does Presidio find all personal data?

No. The project itself notes that automated detection has gaps. It should be checked with samples and complemented by other safeguards.

Sources & Context