Skip to content

Security: Wibias/github-delivery

SECURITY.md

Security Policy

Supported versions

Security fixes are applied to the latest 0.1.x release and to the current main branch.

System and scope

github-delivery is an Agent Skill that reads GitHub evidence, evaluates pull-request policy, reviews code, and performs explicitly authorized GitHub mutations. Security-sensitive components include:

  • natural-language routing and workflow instructions in SKILL.md and references/
  • evidence capture and the authoritative ship gate
  • runtime capability discovery and mutation authorization
  • the installer and deterministic distribution builder
  • GitHub Actions, release artifacts, checksums, SBOMs, and attestations

Threat model and trust boundaries

Treat pull-request content, issue text, review comments, repository files, workflow inputs, external instructions, and downloaded artifacts as untrusted. Important threats include prompt injection that attempts to authorize a mutation, stale-head races, command or path injection, excessive GitHub token permissions, compromised Actions or dependencies, installer path traversal, and release-artifact substitution.

Security invariants

  • Natural-language requests never bypass the authoritative ship gate or mutation policy.
  • GitHub writes require an allowed mutation profile and any required explicit instruction.
  • PR mutations re-check the expected head immediately before execution.
  • Incomplete evidence fails closed and cannot produce a ready verdict.
  • Workflows use least privilege, full-SHA action pins, and disabled checkout credential persistence.
  • Release tags match the package version; artifacts are rebuilt, checksummed, verified, and attested before publication.
  • Install and restore operations remain inside the selected target and backup paths.
  • Secrets, private vulnerability details, and exact human replies are not published without authorization.

Reportable findings and severity context

Report realistic paths to unauthorized GitHub mutations, policy or gate bypass, arbitrary command execution, arbitrary file overwrite, credential disclosure, workflow-token escalation, release tampering, cross-repository confusion, or prompt injection that changes security-sensitive behavior. Severity should reflect reachable impact under the documented trust boundary, not merely the presence of a suspicious string.

Out of scope

The following are not reportable without a concrete security impact in this repository:

  • purely theoretical concerns with no reachable path
  • social engineering that does not exploit skill or workflow behavior
  • denial of service requiring local administrator control of the agent host
  • vulnerabilities in unmodified third-party services or Actions
  • missing optional tools when the workflow already fails closed

Reporting a vulnerability

Do not open a public issue for an undisclosed vulnerability.

Submit reports privately through GitHub's vulnerability reporting form:

https://github.com/Wibias/github-delivery/security/advisories/new

Include:

  • the affected version or commit
  • the impacted workflow or component
  • a realistic abuse path and impact assessment
  • the expected security property
  • minimal reproduction steps or a proof of concept that does not expose third-party systems
  • any suggested remediation, if available

Do not include secrets belonging to third parties, test against systems you do not own, or publicly disclose exploit details before coordinated disclosure.

Response and disclosure timeline

  • We aim to acknowledge a vulnerability report within 7 days.
  • We aim to provide an initial assessment or request additional information within 14 days.
  • We will provide status updates during the investigation when meaningful progress occurs.
  • We aim to resolve confirmed vulnerabilities within 90 days, although complex fixes or coordinated dependency disclosures may require more time.
  • Public disclosure should be coordinated with the maintainer and should normally occur only after a fix or mitigation is available.

If these timelines cannot be met, the reporter will be informed through the private advisory thread.

There aren't any published security advisories