Security policy
Last updated: 28 September 2026
This is the coordinated vulnerability disclosure policy for every Invarix product. It also records how Invarix meets its reporting obligations as a manufacturer under Regulation (EU) 2024/2847, the Cyber Resilience Act.
Report a vulnerability to security@invarix.dk. You will get a human reply within 48 hours.
What is in scope
The published packages: Invarix.Guard, Invarix.Guard.Professional, Invarix.Guard.Evidence, Invarix.Gate and its adapter and bridge packages.
This website and the license issuance flow are in scope for reports about the service itself, such as the checkout or the license email.
Model weight files distributed from the public model releases are in scope for integrity issues, such as a hash mismatch against a published manifest.
How to report
Email security@invarix.dk with the affected package and version, what an attacker can actually do, and the smallest reproduction you have. A failing test, or a policy file plus the tool call that defeats it, is ideal. Tell us if you believe it is already being exploited.
Please do not open a public issue for a suspected vulnerability. If you need encryption, send a first message with no details and we will supply a key.
You may report anonymously. We do not ask you to accept terms before reporting, and we do not ask you to keep quiet about a report once it is fixed.
What you can expect
Acknowledgment from a human within 48 hours. A first assessment with a severity view and a rough plan within 10 working days. A fix or a documented mitigation for confirmed high and critical issues within 90 days of acknowledgment, with the advisory published at the same time.
These are targets from a one-person company rather than contractual commitments, and they are written down here so nobody has to guess. There is no paid bug bounty. Reporters are credited in the advisory unless they ask not to be.
If a fix will take longer than 90 days we will tell you before the 90 days are up, with the reason and a revised date. You are free to disclose publicly after 90 days from acknowledgment whether or not a fix exists. We will not treat good faith research as a license breach and will not pursue legal action over it.
Advisories
Advisories are published on this page and in the changelog of the fixing version. The affected versions are deprecated on nuget.org with a message pointing to the advisory, so Visual Studio and dotnet list package --deprecated warn anyone still on them. Each advisory names the affected versions, the fixed version, the impact, and a workaround where one exists.
INVARIX-2026-001. Approval requests could carry unmasked credentials. Published 27 September 2026. Affects Invarix.Gate, Invarix.Gate.AspNetCore and Invarix.Gate.Evidence 1.0.0-rc.1 and 1.0.0-rc.2. Fixed in 1.0.0-rc.3. No exploitation is known; it was found in an internal review.
When a policy escalates a tool call for human approval, Gate builds a preview of the call's arguments with credentials masked by pattern. In the affected versions that pattern set missed shapes the secret-egress detector recognizes: Azure storage and Service Bus keys, passwords inside PostgreSQL and MongoDB connection URIs under most argument names, and JWTs with an empty or short signature. It also showed part of some credentials it did recognize. A connection string password containing a comma, quote, space or bracket was masked only up to that character, a labeled password kept its first four characters, labeled values inside a JSON document passed as a string were not masked, and a zero width character inside a token could leave the rest of the token shown.
That text appeared in the preview, which Invarix.Gate.AspNetCore posts to the approval webhook and shows on the review page, so escalating a credential egress call could deliver the credential to the webhook destination in the act of asking whether to send it. It could also appear in run reports, the verdict log and the matched fragments written into evidence records. The exposure is limited to destinations the host configures.
Upgrade every Invarix.Gate package to 1.0.0-rc.4, all to the same version. Invarix.Gate.AspNetCore 1.0.0-rc.3 or earlier with Invarix.Gate 1.0.0-rc.4 can fail to send approval requests on .NET 8 and 9. Until you can upgrade, keep secret-egress listed in your policy with its default action, block, so that in enforce mode a credential egress call through the outbound tools it watches is refused without an approval request, and send approval webhooks only to a destination you would trust with the credentials your agents handle.
What is out of scope
Automated scanner output with no demonstrated impact, missing hardening headers with no exploit path, denial of service through deliberately absurd input to a local library, and social engineering of the maintainer.
One recurring non-issue: Gate's free observe mode blocks nothing by design and reports what it would have blocked. That is documented behavior, not a vulnerability. Enforce mode is the enforcing configuration.
Cyber Resilience Act reporting
Invarix is a manufacturer of products with digital elements under Regulation (EU) 2024/2847. From 11 September 2026, for any actively exploited vulnerability in a product above, and for any severe incident affecting the security of those products, we report on these deadlines:
- Early warning within 24 hours of becoming aware.
- Full notification within 72 hours of becoming aware.
- Final report within 14 days of a corrective measure being available, for vulnerabilities.
- Final report within one month of the notification, for severe incidents.
Reports are filed once through the CRA Single Reporting Platform, which routes to the CSIRT designated as coordinator for Denmark, our main establishment, and to ENISA.
This obligation is independent of the disclosure timeline above. A vulnerability that is being exploited is reported to the authorities within 24 hours even where a reporter has asked for a longer embargo, because the regulation requires it. Where the two conflict we tell the reporter which parts were reported and when.
We will also inform affected users, without undue delay and where we have a way to reach them, of an actively exploited vulnerability and any corrective measure they need to apply. For licensed products that is the address used at purchase. The free tier has no such channel, so the advisory and the nuget.org deprecation of the affected versions are the notification.
Security update window
The update window follows the license and is stated there rather than promised open-endedly. A licensed version receives security fixes for the covered window stated in its license, and the last version delivered inside that window keeps working afterwards but stops receiving fixes. Community versions receive security fixes on the current minor version only.
When a version leaves support, that is recorded in the changelog with a date. We do not silently stop maintaining a version.
What the products do and do not guarantee
These are the claims you can hold us to, with the limits that go with them.
All detection and enforcement runs inside your process. No product sends prompt text, tool arguments, or evidence records to Invarix or to any third party. License validation is an offline signature check with no network call.
Gate can only see tool calls routed through an adapter it wraps. A call made on a path that bypasses the adapter is not evaluated. Gate's coverage report exists to make that visible, and it detects such calls where the framework reports them, but it cannot prevent them.
Tools that execute inside a model provider's own service cannot be intercepted before they run. Agents hosted in Azure AI Foundry Agent Service or Copilot Studio are outside the enforcement boundary entirely.
Enforce mode fails closed. A detector that throws denies the call by default, and that default is configurable and visible.
Evidence records are hash chained and signed, and proofs verify offline against a published public key. That makes them cryptographically verifiable and admissible. It does not confer the legal presumption of integrity that under eIDAS belongs only to a ledger operated by a qualified trust service provider.
Contact
security@invarix.dk for vulnerabilities. sales@invarix.dk for everything else.