This policy explains how to report vulnerabilities in Eneo and how the maintainers triage security work. Do not report sensitive vulnerabilities in public GitHub issues, pull requests, or discussions.
Report a vulnerability privately in one of two ways:
- On GitHub, through Security → Report a vulnerability (https://github.com/eneo-ai/eneo/security/advisories/new). The report is visible only to you and the maintainers, and the fix can be coordinated in a private advisory.
- By email to
security@sundsvall.se.
Include as much of the following as possible:
- affected version, commit, branch, or deployment environment
- affected component, endpoint, package, workflow, or Docker image
- impact and the data, tenant, credential, or permission boundary involved
- reproduction steps or a minimal proof of concept
- relevant logs, screenshots, request IDs, or dependency/advisory links
- whether the issue is already public or under active exploitation
- suggested remediation, if known
- contact details for coordinated follow-up
Please do not publicly disclose the issue until a fix or mitigation is available and maintainers have had a reasonable coordination window.
These are targets, not guarantees. Active exploitation, exposed secrets, malware, or vulnerabilities affecting authentication, authorization, file parsing, prompt handling, LLM provider credentials, SSRF, RCE, or deserialization take priority.
| Severity | First response | Initial triage | Target fix or mitigation |
|---|---|---|---|
| Critical | 1 business day | 1 business day | As soon as practical |
| High | 1 business day | 2 business days | 30 days |
| Medium | 2 business days | 5 business days | 90 days |
| Low | 5 business days | Next maintenance cycle | Best effort |
Security fixes land on develop first. Supported release branches are any
release branches that are currently deployed or explicitly maintained by the
Eneo maintainers. Older tags and inactive branches are not guaranteed to receive
security fixes.
If a fix affects users outside the hosted/deployed environments, maintainers should publish upgrade or mitigation guidance through a GitHub Security Advisory.
The repository uses GitHub security features and CI to prevent regressions:
- Dependabot creates weekly dependency update pull requests for backend, frontend, GitHub Actions, Dockerfiles, and devcontainer configuration.
- Dependency Review runs on pull requests into
developand blocks newly introduced vulnerable dependencies athighseverity or above. - The reusable
Securityworkflow audits the complete runtime Python lock and frontend lock, runs Semgrep security-audit/GitHub Actions rules, and scans the added Git history with Gitleaks. It is included in the required CI aggregate and runs before image publication. The dependency gate blockshighandcriticalfindings, including those without a fix; lower severities emit warnings. Exact, unexpired exceptions are retained. Unknown severities, failed scans, invalid reports and expired exceptions still fail the check. Python severity ratings come from the GitHub Advisory Database and are retained alongside the raw audit reports as CI artifacts. - Temporary dependency mitigations belong in
frontend/patches/SECURITY.md, with an owner, exact advisory, regression tests and review expiry. The audit retains the raw finding and only accepts the named mitigation after its tests pass. - Gitleaks exclusions in
.gitleaks.tomlrequire both an exact file path and an exact reviewed test credential or generated schema example. They apply only to the named rule, survive squash commits and line moves, and do not exclude entire files, directories or rules. The security workflow tests that changed values and values outside their reviewed files are still detected. Release review includes a full-history audit and confirmation that any real historical credentials have been revoked; a diff scan cannot prove revocation. - CodeQL advanced setup is paused while maintainers evaluate a code-scanning setup that gives developers useful feedback without high false-positive noise. Re-enable code scanning only after the query set and triage process are tuned enough that alerts are actionable for maintainers.
- Release SBOMs are generated from the published backend (including the React web build), Svelte frontend, tools runtime, and bundled
SeaweedFS container
image digests and attached to GitHub Releases as CycloneDX JSON, SPDX JSON,
and human-readable table files. The backend release also includes a narrower
CycloneDX SBOM for the Python runtime installed in
/app/.venvinside the released backend image. The frontend release also includes a source dependency SBOM generated with CycloneDX Generator after installing the frontend from the checked-out release tag with the frozen Bun lockfile, so JavaScript dependencies remain visible even when the runtime image only contains the bundled build output. Release SBOM assets are attested through GitHub artifact attestations, and the pushed backend, frontend, and SeaweedFS images include provenance attestations in GHCR.SBOM-SHA256SUMS.txtcovers the SBOM files;IMAGE-DIGESTS.txtlabels the exact image digests scanned by platform and records the bundled store's multi-platform manifest reference for deployment. - The bundled object store is an Eneo-built artifact, not an upstream-signed
SeaweedFS image.
.github/workflows/seaweedfs_image.ymlowns it and runs only when its pinned inputs change, so application image publishing is never blocked by object-store policy drift. Privileged build, attestation, and publication jobs run only fordevelopand use the GitHub OIDC workflow identity—no PAT, registry password, or signing key. Pull requests and manual runs from other branches run the same source policy, build, smoke test, SBOM, and vulnerability gate against an ephemeral local registry with read-only permissions; nothing reaches GHCR and nothing is attested. The workflow pins the SeaweedFS 4.40 source commit/archive, Dockerfile frontend, compiler image, runtime image, tools, and every Action by immutable digest or full commit SHA. A hashed downstream patch upgrades dependencies with known advisories until upstream includes those fixes. It publishes only after both Linux platform candidates pass source/license policy, a reproducibility rebuild, the exact reference Compose bootstrap and persistence smoke, a reachable Go vulnerability scan, and a Grype gate that includes unfixed HIGH/CRITICAL findings. - The published tag is
ghcr.io/eneo-ai/eneo-seaweedfs:<docker/seaweedfs/VERSION>and is immutable by policy: the serialized publish job refuses to move a version tag to different content and fails closed when it cannot read the registry. An identical rebuild leaves the tag untouched and emits fresh attestations bound to its own source commit. Every change to the image inputs bumpsVERSION. A weekly scheduled run re-audits the pinned source and both published platform digests against current Go and Grype vulnerability databases without rebuilding or publishing; a red scheduled run is the signal to bump the downstream patch (seedocker/seaweedfs/README.md). - SeaweedFS source-license classifier gaps are resolved only by exact
module/version/license-file hashes in
docker/seaweedfs/verify-supply-chain.sh. Any dependency, version, license content, or unknown-package drift fails the job for review. This is not a general allowlist or vulnerability waiver. - Configure the
object-content-supply-chainGitHub Environment as a protected trusted-publication boundary with a deployment branch policy that allows onlydevelop; the workflow enforces the same rule in code, and the environment policy is the backstop. Do not grant package write, OIDC, or attestation permissions to pull-request workflows, and do not make the path-filtered SeaweedFS workflow a required status check. The canonical image/SBOM owners remain.github/workflows/build_and_push_images.yml,.github/workflows/seaweedfs_image.yml, and.github/workflows/release_sbom.yml; do not create a parallel signing system. - Secret scanning and push protection should remain enabled for provider keys, tokens, credentials, and other repository secrets.
- The normal
CIgate validates frozen backend and frontend installs before dependency update pull requests are merged.
The develop ruleset requires CI and Dependency Review, and the PR label
classification ruleset requires PR labels on develop, main and
release/**. While code scanning is paused, remove CodeQL gate from the
GitHub ruleset; otherwise pull requests can wait for a check that no longer
runs.
Access to the eneo-ai organisation follows least privilege. Core Maintainers
(the Eneo team at Sundsvalls kommun) steward the platform and are accountable
for what is merged and released. Collaborators from partner municipalities and
agencies contribute through pull requests. The contributor-facing explanation,
including the role table, diagrams and an FAQ, is
Governance and Roles
on the documentation site.
- Roles.
@eneo-ai/core-maintainersis the only team with write-level (Maintain) access. It also holds the organisation's Security manager and CI/CD admin roles. Collaborators (@eneo-ai/collaborators) have Triage access and contribute from forks. A small number of organisation owners hold Admin rights for rulesets, team access and repository settings. The organisation base permission is read. develop. Changes arrive only through pull requests. At least one approval from@eneo-ai/core-maintainersis required, enforced by the ruleset and requested automatically through.github/CODEOWNERS. New commits dismiss earlier approvals, review conversations must be resolved, and merges go through the merge queue.- Releases. Only Core Maintainers can create, update or delete
release/**branches andv*tags. Av*tag starts image publication, so this restriction controls who can publish a release. main. Protected against deletion and force pushes.- Bypass. Core Maintainers can bypass the
developruleset for a single pull request when a fix is urgent or no second Core Maintainer can review. A bypass also skips the merge queue and required checks, so use it only when every check has passed on the pull request, record the reason in the pull request, and have another Core Maintainer review the change afterwards. Repository administrators are not exempt from rulesets unless they are on the bypass list. - Workflow trust. Pull requests from forks run with a read-only
GITHUB_TOKENand without repository secrets. Workflows started from branches in this repository can request package write permissions, which is why write access is not granted outside Core Maintainers. Keep repository and organisation secrets limited to those referenced by workflows; when a secret is no longer used, delete it and revoke it at its source. - Changes to access. Only organisation owners change team roles, rulesets or bypass lists. Record each change in a GitHub issue so the access history is reviewable.
Use this order for existing Dependabot, malware, secret scanning, Dependency Review, Docker, GitHub Actions, and any enabled code-scanning alerts.
- Dependabot malware alerts
- secret scanning alerts
- critical runtime vulnerabilities
- issues affecting authentication, authorization, file parsing, prompt or LLM provider credentials, SSRF, RCE, or deserialization
For Eneo, treat packages such as LiteLLM, LangChain, MCP clients/servers, HTTP clients, cryptography libraries, FastAPI, document parsing libraries, frontend frameworks, and build tooling as higher risk because they process prompts, files, network calls, credentials, or user-controlled content.
- high severity runtime vulnerabilities
- vulnerable Docker base images used in production
- vulnerable GitHub Actions used in trusted workflows
- medium runtime vulnerabilities
- critical or high dev-only vulnerabilities that execute during build or test
- low severity alerts
- dev-only issues with no practical exploit path
- major version update backlog
Do not dismiss alerts just to reduce the count. Dismiss only with a clear reason such as "vulnerable code not used", "false positive", or "tolerable risk", and record an expiry date or follow-up issue.
For each new alert:
- Classify it as malware, secret exposure, runtime dependency, dev dependency, Docker image, GitHub Action, or code scanning.
- Check whether it is reachable in Eneo and whether it touches user input, files, prompts, credentials, authentication, network calls, or deployment.
- Prefer the Dependabot security PR. If Dependabot cannot create one, manually bump, pin, replace, or remove the dependency.
- Review the diff. For AI and LLM packages, check changelogs for telemetry, callback behavior, environment variable handling, proxy behavior, logging changes, and changed defaults.
- Run frozen-lockfile installs, CI, and any relevant smoke tests.
- Merge only after the relevant checks pass or an explicit risk exception is documented.
- Verify the alert closes. If it does not, inspect lockfiles for remaining vulnerable transitive dependencies.
- If the issue cannot be fixed immediately, create a GitHub issue or security tracking item with owner, risk, mitigation, and target date.
Treat incidents in packages that handle LLM provider keys, prompts, proxying, model routing, callbacks, or request logging as higher severity than ordinary library updates.
Immediate response:
- check whether Eneo uses the affected package and version
- use the Dependabot security PR if GitHub has an advisory
- manually bump, pin, or temporarily remove the package if no advisory exists yet
- rotate any potentially exposed provider keys
- review application logs, CI artifacts, Sentry events, build logs, and package install logs
- publish a GitHub Security Advisory if users need to upgrade or take action
- add a regression or security test when the incident involved logging, credential handling, SSRF, authentication bypass, or unsafe defaults
Dependabot is useful once an advisory exists, but it is not a replacement for incident response when the ecosystem learns about a compromise before the GitHub Advisory Database is updated.
Use GitHub Actions secrets for CI and deployment credentials. Use Dependabot secrets only when Dependabot needs access to a private package registry. Do not store runtime application secrets in Dependabot secrets.
If a secret is exposed:
- revoke or rotate it
- inspect logs and artifacts for use or exfiltration
- remove it from repository history if needed
- add or adjust secret scanning patterns when the secret format is specific to Eneo
Eneo's structured-log path enforces redaction on both stdout log attributes and auto-instrumented span attributes:
Authorization,Cookie, andSet-Cookieheaders are never logged.- Query parameters named
code,state,token,access_token,refresh_token,client_secret, or matching*token*or*secret*(case-insensitive) are replaced with[REDACTED]in both log lines and span URL attributes. - Request and response bodies are not logged by default; adding body logging requires a deliberate code change and a redaction review.
user_emailis logged as a deliberate carryover from existing audit behavior. It is high-cardinality PII; configure the log aggregation system not to pre-index it.
The redaction policy is unit-tested across both layers (stdout logs and span attributes from FastAPI, httpx, and aiohttp instrumentation). For the full policy and verification commands, see OBSERVABILITY.md §8.
Track security work in GitHub issues or GitHub Security Advisories, not in
private chat threads. Use labels such as security, dependencies,
backend, frontend, docker, and github-actions so maintainers can filter
the backlog. Add unresolved security issues to the Eneo project board and set
the Security status when the item needs team visibility.
Each accepted risk or delayed fix needs:
- owner
- affected component
- reason for delay
- mitigation
- target date
- review date or expiry condition
Object-content image policy has no standing exceptions. A future CVE or license exception must name the exact advisory/license and artifact digest, owner, reason, compensating control, and expiry. Absent that human-approved record, the publication fails closed.
- Use frozen installs for dependency changes.
- Do not bypass security gates without a documented exception.
- Keep dependency PRs small enough to review.
- Read changelogs for packages that handle prompts, credentials, networking, file parsing, authentication, authorization, or deployment.
- Prefer patch and minor updates first while the alert backlog is being reduced.
- Revisit the Dependency Review threshold after the critical and high backlog is stable.