Skip to content

Security: defilantech/meitheal

Security

SECURITY.md

Security Policy

Supported Versions

Version Supported
0.1.x ✅

Reporting a Vulnerability

We take security seriously. If you discover a security vulnerability in Meitheal, please report it responsibly.

How to Report

  1. Do NOT open a public GitHub issue for security vulnerabilities.
  2. Preferred: use GitHub's private vulnerability reporting. Reports stay private, are tracked in one place, and let us publish an advisory once a fix ships.
  3. Alternative: email contact@defilan.com.

What to Include

  • Description of the vulnerability
  • Steps to reproduce
  • Potential impact
  • Suggested fix (if any)

Response Timeline

  • Acknowledgment: Within 48 hours
  • Initial Assessment: Within 7 days
  • Resolution Target: Within 30 days for critical issues

Scope

This security policy applies to:

  • The meitheal-operator
  • The meitheal-agent (in-cluster on Linux, and native on Linux and macOS hosts)
  • The meitheal CLI
  • The Helm chart
  • Container images published to GHCR

Out of Scope

  • Third-party dependencies (report to upstream)
  • LLM model vulnerabilities (report to model providers)
  • The model servers Agents call, including LLMKube (report to those projects)

Trust model

Meitheal turns custom resources into code changes made by a model with tools, so write access to its resources is powerful. Grant it the way you would grant access to the credentials and hosts the agents use.

  • AgenticTasks and Workloads. Creating one makes an Agent clone a repository, run a model-driven loop whose tools include a shell in the task workspace, and push branches with the git credentials the meitheal-agent was given. In the default InProcess execution mode that loop runs in the meitheal-agent itself (its pod, or the Linux or macOS host it runs on natively); in Job mode it runs in a coder Job pod. A task names its Agent with spec.agentRef, which resolves in the task's own namespace.
  • Agents. An Agent's spec.execution.image and spec.execution.serviceAccountName become the image and ServiceAccount of the coder Jobs the meitheal-agent submits, so write access to Agents is equivalent to creating pods with that ServiceAccount in the namespace those Jobs run in. A cloud-proxy Agent names an endpoint (spec.providerConfig.baseURL) and a Secret key (spec.providerConfig.apiKeySecretRef, resolved in the task's namespace); the agent sends that Secret's value as a bearer token to that endpoint with every model request. Grant create, update and patch on Agents only to principals you would also trust with those Secrets.
  • Credentials come from Secrets. The meitheal-agent reads its GitHub token from the Secret named by the chart's agent.githubToken.secretName, and coder Jobs get theirs from coder.gitCredentialsSecret. The chart's meitheal-agent ClusterRole and the coder Job Role both grant get on Secrets, without a resourceNames restriction.
  • Cloud providers are a kill switch, not a default. The operator's --allow-cloud-providers flag (chart value meitheal.allowCloudProviders) defaults to true. Set it to false on air-gapped or compliance-restricted clusters: the Workload controller then drops every step whose Agent uses a non-local provider and reports a CloudReviewersSuppressed condition. It does not apply to an AgenticTask created outside a Workload.

Verifying releases

Release images are signed with cosign and carry SLSA build provenance and SBOM attestations, and the .tar.gz release archives carry SLSA build provenance, all made by the release-please.yml workflow on main. The darwin binaries are code-signed and, in releases built with the signing secrets, notarized by Apple. Verifying Meitheal releases has the commands that check each of these.

Security Best Practices

When deploying Meitheal:

  1. Use RBAC: Restrict who can create Agents, AgenticTasks and Workloads (see Trust model)
  2. Scope credentials: Give the agent a GitHub token limited to the repositories it works on, and keep model API keys in their own Secrets
  3. Network Policies: Enable the chart's networkPolicies and isolate the agents from workloads they do not need to reach
  4. Cloud providers: Set meitheal.allowCloudProviders: false where no request may leave the cluster
  5. Resource Limits: Always set CPU/memory limits to prevent DoS
  6. Image Verification: Verify release images before you deploy them (see Verifying releases), and use image digests in production Helm values

There aren't any published security advisories