Skip to content

Security: obentoo/bentoolkit

SECURITY.md

Security Policy

Supported Versions

bentoolkit is pre-1.0 and ships frequent releases. Security fixes land on the latest released minor only; please upgrade before reporting.

Version Supported
0.42.x ✅
< 0.42 ❌

Reporting a Vulnerability

Please do not report security vulnerabilities through public GitHub issues, pull requests, or discussions.

Report privately through GitHub's Private Vulnerability Reporting:

  1. Open the new advisory form.
  2. Click Report a vulnerability and fill in the details.

This keeps the report confidential, lets us collaborate on a fix in a private fork, and supports CVE issuance when warranted.

Please include:

  • A description of the vulnerability and its impact.
  • Steps to reproduce (a minimal proof of concept is ideal).
  • Affected version(s), operating system, and Go toolchain.
  • Any suggested remediation, if you have one.

What to expect

  • Acknowledgment within 3 business days.
  • Initial assessment within 10 business days, with a severity estimate and an expected timeline.
  • Progress updates at least every 10 business days until resolution.

This is a community-maintained project, so timelines are best-effort.

Coordinated Disclosure

We follow coordinated disclosure. Please give us a reasonable window to ship a fix before any public disclosure — by default up to 90 days from acknowledgment, or sooner once a fix is released. We are glad to credit reporters in the advisory unless you prefer to remain anonymous.

Scope

In scope — the bentoolkit Go codebase (the overlay and snapshot modules), its build chain, and the CI/release workflows in this repository.

Out of scope — vulnerabilities in upstream software that bentoolkit merely orchestrates or depends on (Gentoo/Portage, third-party overlays, btrbk, snapper, systemd, btrfs); please report those to their respective projects. Because snapshot and overlay operations run with elevated privileges by design, issues that require pre-existing root/privileged access are out of scope unless they cross a clear privilege boundary.

Security Measures

Every change to this repository is gated in CI by:

  • govulncheck — reachability-based scanning against the Go vulnerability database (pinned via a go.mod tool directive).
  • OSV-Scanner — dependency CVE scanning.
  • gitleaks — secret scanning across the full git history.
  • zizmor — GitHub Actions workflow hardening; every action is pinned by commit SHA.
  • Dependabot — weekly dependency updates with a 7-day release cooldown to dodge the window when most hijacked or yanked packages are caught.

You can reproduce the dependency audit locally with make audit.

Secret Handling

bentoolkit resolves secrets at runtime instead of storing them in its configuration files. Neither config.yaml nor snapshot.toml holds a token, API key, or password at all. Every secret bentoo consumes — the GitHub token, per-repository tokens, the LLM API key, the authenticated-fetch serial, the ntfy token, and the SMTP password — is resolved through a single chain:

  1. an environment variable, then
  2. the user secrets file $XDG_CONFIG_HOME/bentoo/secrets (else ~/.config/bentoo/secrets), then
  3. the system secrets file /etc/bentoo/secrets.

The secrets file is .env style (NAME=value, # comments, an optional export prefix) and should be chmod 600; bentoo warns once if it is group- or world-readable.

What a secret value can and cannot reach:

  • Secret values never appear in logs, in argv, or in error messages.
  • An LLM agent — every claude process bentoo spawns — receives an allow-listed environment and nothing else: PATH, HOME, TMPDIR, LANG, TERM, CLAUDE_CONFIG_DIR, NODE_EXTRA_CA_CERTS, SSL_CERT_FILE, HTTP_PROXY, HTTPS_PROXY, NO_PROXY and their lowercase forms, and every LC_* and XDG_* variable. In bare mode it also receives exactly one ANTHROPIC_API_KEY, holding the key bentoo resolved; otherwise it receives no API key at all and uses the CLI's logged-in session. The manifest fixer additionally receives PORTAGE_* and one DISTDIR, the scratch directory bentoo computed for it. GITHUB_TOKEN, the ntfy token, the SMTP password and every BENTOO_* variable stay out.
  • Other subprocesses — pkgdev, git and ebuild run by bentoo itself — still inherit bentoo's environment, so a secret exported in the shell that starts bentoo reaches them. The build-validation child (ebuild for the compile gate) is the exception: it gets its own allow-list. Keep secrets in the secrets file rather than in the environment where you can.

A config file that still carries a secret key removed in an earlier release gets one actionable warning naming the key and where the value belongs, emitted when the file is loaded — a stale key is never ignored in silence.

LLM Agents

Some autoupdate operations hand a task to an LLM agent (the claude CLI) that reads untrusted upstream pages. Each agent is started with a fixed permission set; no configuration key widens an agent's tools, hosts or paths, and a refused tool ends the run as a normal failure that names the tool (never the refused call's input) instead of being retried with more.

Agent Tools Files reachable Network
Text client (version extraction, reviews) none none; runs in a private 0700 directory created per call and removed after none
Manifest fixer Read, Edit, Write, Bash(pkgdev *), WebFetch the package directory WebFetch to the package's upstream hosts
Registry fixer Read, Edit, Write, WebFetch the .autoupdate config directory WebFetch to the registry entry's hosts
Build fixer Read, Edit the staged package directory none
Bump reviewer Read a private 0700 directory created per review and removed after none
  • Directory scope. Read is granted as Read(//<dir>/**) and writes as Edit(//<dir>/**) (the CLI applies Edit rules to every file-writing tool). A directory containing anything outside [A-Za-z0-9._+@/-] — a space, for example — is refused before the agent is spawned.
  • Secrets files are denied by path. Every agent is denied Read of $XDG_CONFIG_HOME/bentoo/secrets (else ~/.config/bentoo/secrets) and of /etc/bentoo/secrets; an agent that can edit is also denied Edit of both.
  • WebFetch host rule. WebFetch is granted only as WebFetch(domain:<host>) for the package's upstream hosts — the hosts of its registry url, mirrors and fallback_url, and, for the manifest fixer, of the http(s) URLs pkgdev printed — plus github.com, codeload.github.com and objects.githubusercontent.com. A host that is not a lowercase DNS name is dropped with a warning, so a value in packages.toml or in pkgdev output cannot widen a rule or forge a second one. So is an IPv4 literal — any host whose last label is a number (127.0.0.1, 169.254.169.254, 127.1, 127.0.0.0x1). The check reads how a host is spelled, not where it resolves: a DNS name that points at loopback, a private network or a metadata endpoint (for example 169.254.169.254.nip.io) is still granted if it appears in a URL upstream printed. No agent holds curl, wget, cat or ls.
  • Pinned settings. Every agent runs with --permission-mode dontAsk (a call that is not pre-approved is refused, never prompted), loads no user, project or local settings — so neither ~/.claude/settings.json nor an overlay's .claude/settings.json can widen it — loads no MCP servers or account connectors (--strict-mcp-config), and runs with inline settings that set permissions.blockReadsOutsideWorkingDirectories to true and permissions.disableBypassPermissionsMode to "disable".
  • Residual risk. Permission rules govern the agent's own tools, not the programs those tools run. pkgdev, which the manifest fixer may run, reads files and fetches distfiles by itself, outside these rules. bentoo applies no OS-level sandbox to the agent; run it under one (a dedicated user, firejail, a container) if that residual risk matters for your host. The same control is the only one that closes the DNS-alias case above: a name such as 169.254.169.254.nip.io passes the host rule, and resolving it when the rule is built would not help, because the name can resolve differently by the time the agent fetches. Egress rules for the user the agent runs as (refusing loopback, RFC 1918 and link-local destinations) are what stop it.

There aren't any published security advisories