Security
Security, plainly.
What Penweir does with your data, what it never does, and where its limits are. Penweir is in development: this page describes the design we are building to. We will publish verifiable details (code, SBOM, test results) with the first release.
Architecture
Everything happens inside your network.
Penweir runs next to your systems, on your hardware or your cloud account. Requests come in from your applications, pass through the pipeline in memory, and leave only towards the model your policy allows.
Data handling
What is kept, and what is not.
| Data | Where it lives | How long |
|---|---|---|
| Request payloads (the actual content) | Process memory only | Gone when the response returns. Never written to disk or logs. |
| Audit records | Your database | Your retention policy. Metadata only: time, pipeline, source, field counts, model, latency, status. |
| Pipeline and policy configuration | YAML files in your repository | Versioned by you, reviewable in pull requests. |
| Licence key | Your deployment | Checked offline. Penweir never contacts us. |
| Anything sent to us | Nowhere | There is no telemetry and no licence server. |
Redaction
Designed to fail safely.
No detector finds every name, address or account number. Pattern matching misses unusual formats; language models miss context. We would rather say that here than have you find out. So Penweir does not rely on detection alone:
Allowlist first
A pipeline names the fields that may leave. Everything else is dropped before any detector runs, so an unknown field never reaches a model by accident.
Fail closed
If a step errors, times out or cannot classify a request, the request is held, not passed. Held requests appear in the audit log.
Route instead of risk
Requests classified as sensitive can go to a model you run yourself, so the data never leaves your network at all.
Typed, irreversible tokens
Redacted values become tokens like [EMAIL]. The original is not kept anywhere in Penweir, so it cannot leak later.
How each detector works, and how it performed in our tests, will be documented with the first release.
Compliance
What Penweir helps with, and what it doesn't.
GDPR / AVG
Penweir helps you apply data minimisation (art. 5(1)(c)): models only see the fields they need. Because it runs on your infrastructure, we never become a processor of the data that flows through it.
Records of AI use
The metadata-only audit log shows which kind of data reached which model, and when. Useful evidence for your own records under the GDPR and the EU AI Act.
Not a certificate
Penweir is a control, not a compliance programme. It does not make an organisation compliant on its own, and we don't claim certifications we don't have.
Releases & support
How updates will work.
Semantic versioning
Breaking changes only in major versions, always with upgrade notes.
Signed releases
Every release package will be signed and ship with a software bill of materials (SBOM).
Security fixes
During your update period, critical vulnerabilities get a fix or documented mitigation within 7 days of confirmation, and every licence holder is told by email.
No kill switch
When your update period ends, your version keeps running. No forced upgrades, no remote shutdown.
Vulnerability disclosure
Found something? Tell us.
Email the details to the address below. Please give us a chance to fix the issue before you publish it.
security@penweir.com- We confirm receipt within 2 business days.
- We keep you informed while we work on a fix, and credit you if you want.
- We follow coordinated disclosure, with a default window of 90 days.
- We will not take legal action against good-faith research that respects this policy.