Skip to main content
Version: 1.3

Security process

This page describes how DeviceChain handles vulnerability reports, how it meets the reporting duties of the EU Cyber Resilience Act, what it publishes about each release, and how the AI agents that help maintain the project are kept away from anything that could ship a compromised build. It says what is in place today and what is not yet.

Reporting a vulnerability​

Report privately, never in a public issue. The routes, and the supported versions, are in SECURITY.md. A machine-readable contact is published at /.well-known/security.txt.

StepTarget
Acknowledge and triage: read it, reproduce or reject it, make a first severity callwithin 24 hours, every day of the week
Full assessment: valid, severity, affected versionswithin 7 days
Fix, by severityCritical 14 days, High 30, Medium 90, Low the next release

These are targets held by a small team, not guarantees. If one is going to be missed, the reporter is told why.

Reporting duties under the Cyber Resilience Act​

Since 2026-09-11 a manufacturer must report to the authorities a vulnerability in its product that is being actively exploited, and a severe incident that affects the security of the product. Both follow the same three steps, counted from the moment the project becomes aware:

StepActively exploited vulnerabilitySevere incident
Early warningwithin 24 hourswithin 24 hours
Notification, with what is known so farwithin 72 hourswithin 72 hours
Final report14 days after a corrective or mitigating measure is available1 month after the notification; if the incident is still ongoing then, a progress report at that point and the final report within 1 month of handling it

The runbook the maintainers follow:

  1. Decide whether the clock has started. Active exploitation means credible evidence that someone used the weakness against a real system without the owner's permission. A proof of concept sent by a researcher does not start it. A report of use in the wild does.
  2. Send the early warning through the single reporting platform the EU agency for cybersecurity operates, which routes it to the national coordinator (CSIRT) of the member states concerned. It states that the report is an early warning, whether malicious action is suspected, and the member states where the product is available.
  3. Send the notification within 72 hours: the nature of the vulnerability, the affected versions, any mitigation already available, and an assessment of severity.
  4. Tell users without undue delay, as soon as the early warning is out and before a fix exists: what is affected and any mitigation they can apply now (a GitHub Security Advisory draft, a release note, or a notice on the repository).
  5. Ship the fix. A fixed patch release and the published GitHub Security Advisory (with a CVE where one applies) go out together, and the advisory says what to do before upgrading if there is anything.
  6. Send the final report with the root cause, the fix, the malicious actor if known, and how users were informed.
  7. Record the dates of each step in the private advisory so the next incident starts from a working example.

DeviceChain is open-source software that is not sold, so whether a given deployment makes its operator the manufacturer for these purposes is for that operator to determine. The project runs this process for the software it publishes regardless.

What each release publishes​

Every release carries a signed SBOM and build provenance for each image (Verifying releases). When an advisory affects a dependency, the maintainers check the release SBOMs for the affected package and version, list the images that are exposed, and publish a statement in the OpenVEX format with the advisory: affected or fixed for the images that carry the package, and not affected with a stated reason for those that ship it but cannot reach the vulnerable code.

The check is a script in the repository, hack/vex-exposure.sh. It reads an SPDX SBOM per image and an advisory, lists exposed images, and drafts the statement. It treats an SBOM with no packages as an error, never as "nothing exposed".

Reproducible builds​

A build is reproducible when building the same commit twice, on different machines and in different directories, produces byte-identical output. A job in CI builds dcctl and the edge agent from two unpacked copies of the commit with separate caches and fails if the digests differ. It is exercised by a planted nondeterministic build, which it must catch. Container images and the Helm chart are not yet covered by that job; the broker image has its own reproducibility check.

AI-assisted development​

AI agents draft code, documentation and reviews for this project. They are treated as an untrusted contributor with an unusually fast keyboard, and the controls do not depend on the agent behaving well:

  • No secrets. Agents run without access to signing identities, registry credentials or cloud keys. The release workflow signs with a short-lived identity bound to the repository, and no long-lived signing key exists.
  • No release authority. An agent cannot cut a tag or publish an artifact. The release workflow runs only from a maintainer-created tag.
  • No merge to the protected branch. An agent opens a pull request. A maintainer merges it after the required checks pass, the same as for any contributor.
  • Untrusted input is data. Issue text, pull-request descriptions, advisory reports, web pages and dependency documentation that an agent reads are treated as data to analyse, never as instructions to follow. A report that says "ignore your rules" is a finding.
  • Review. A change an agent wrote is reviewed by a maintainer and, for risky areas, by a second independent reviewer, and a new guard is shown to fail on a planted defect before it is trusted.

Not yet in place​

  • Reproducibility checks for container images and the Helm chart.
  • Replacing the remaining long-lived cloud credential used for performance testing with short-lived, workload-identity access from CI.
  • A published target score against the OpenSSF Scorecard and the OpenSSF baseline controls. Scorecard already runs weekly and its result is public.