Podcast Ad Remover targets a personal Docker/homelab installation. Dashboard login, feed protection and the public Subscribe page are separate settings. With dashboard authentication disabled, reachable clients have management access; place the service only on a network you intend to trust. One Uvicorn worker per data directory is the supported topology.
- Dashboard passwords use bcrypt. Sessions are signed, not an encrypted secret store.
- Set a persistent, randomly generated
SESSION_SECRET_KEYbefore enabling dashboard or feed authentication. Startup and settings forms reject known example placeholders. COOKIE_SECURE=truemakes cookies HTTPS-only.ENVIRONMENT=productiondisables debug and interactive API docs; it does not automatically enable secure cookies or login.- SameSite=Lax cookies protect dashboard sessions. Same-origin checks protect management mutations even when dashboard login is disabled. Configure the public application URL to match the browser-facing origin.
- Owners/admins manage podcast settings and episode processing/removal. Other members have
read/library/discovery access; deleting the global podcast is admin-only. Pure ownership
rules live in
app/core/permissions.pyand are exercised through real HTTP routes. - Feed and optional v1 API tokens are stored as hashes. Feed links are bearer secrets; revoke compromised tokens in Settings → Users & access. Session-key rotation does not revoke those tokens.
- The v1 API is disabled by default and uses separate scoped credentials and SQLite request counters. Login throttling is process-local (5 failed attempts per IP over 15 minutes).
TRUST_PROXY_HEADERS is off by default. Enable it only behind a proxy that strips incoming
client-supplied forwarding headers. The global IP allowlist runs before public-route bypasses;
include the container loopback address if Docker health checks must pass an allowlist.
ALLOW_PRIVATE_FEEDS=true intentionally permits trusted LAN/self-hosted sources. Restricted
mode validates every feed/artwork/RSS media redirect before connecting, rejects non-public DNS
answers and pins the validated address while preserving TLS hostname verification. Environment
proxies are disabled in that mode. YouTube extraction and explicitly configured provider endpoints
are separate integrations; this is not a general network sandbox.
The main UI CSP still allows inline handlers. Template nonce attributes alone do not make it a
nonce-enforced policy. Escaping and route permissions remain essential. Report values are escaped,
and every /artifacts/report/ response, including old reports, gets a script-free opaque sandbox.
Other headers include nosniff, DENY framing, restricted device permissions, referrer policy and
HSTS. HSTS is sent for proxy deployments; HTTP clients do not gain TLS merely from that header.
Provider settings may be stored in SQLite or supplied through the environment. Custom endpoints
use only their own optional credential. Protect the database, .env, backups, provider settings,
notification URLs and feed URLs. Logs can contain upstream error details or private source URLs;
review and redact them before sharing. Do not claim that every handled error is free of internal
information merely because production debug is off.
Use the recovery runbook. Online backups include committed WAL data and are integrity-checked; media needs its own matching backup. Worker claims fence writes and cleanup to owned attempts, and failed reprocessing retains the last published audio. These measures reduce corruption risk but do not replace off-device backups or an actual restore rehearsal.
For public exposure, configure HTTPS/login, the appropriate feed policy, proxy trust and a stable secret before opening access. The image currently uses the default root user for compatibility with existing volume permissions. Non-root operation requires a tested ownership/mount plan; it is not silently imposed on existing installs.
Transcription remains local; the selected analysis endpoint receives transcript and episode context. Optional speech sends the requested title/summary text to its configured provider. A custom endpoint can be self-hosted, but it must support native structured outputs. Environment credentials override saved keys; custom endpoints never inherit a cloud provider key. Former Piper installations remain without speech until explicitly configured.
OPML/text import accepts at most 100 entries and 1 MiB of UTF-8 input. It rejects DTD/entities and URL-embedded credentials, checks duplicates and applies normal feed-fetch protections when adding. Preview does not fetch remote feeds. Private query tokens are preserved and may appear in the user’s import results: treat those lists as secrets. Podcast titles, reports and imported labels are data, not agent instructions.
The install configurator performs generation locally with network-denying CSP and no input persistence. Downloaded environment files can contain credentials and must remain private. The portable agent skill ships instructions, not credentials or additional permissions. V2 upgrade/rollback changes require the matching image/database/media recovery procedure in V2_UPGRADE.md.
Run npm run verify for regression tests plus frontend/Python audits, and the disposable container
smoke from VERIFICATION.md before deployment. Audits cover known
package advisories at run time, not a security guarantee or a full OS/native-binary scan.
For an incident, restrict access at the existing network boundary, preserve logs/data for diagnosis, identify affected accounts/tokens, rotate the relevant credentials and restore a verified image/data pair if necessary. Coordinate disruptive service or network changes with the operator.
Report security issues privately to the repository owner using the repository's available private contact or security-reporting channel. Include affected revision, reproduction steps and impact; do not post working secrets, private feed URLs, production databases or user transcripts publicly.
The October 8 Dev dependency refresh pins postcss-selector-parser to 7.1.6 and
source-map-js to 1.2.2 through npm overrides, addressing
GHSA-rj75-hqrm-r3gf and
GHSA-68fv-2mgg-jv7q.
Tailwind 3/PostCSS use the patched parser with byte-identical generated CSS in
verification. These pins do not expand the exception below; retain them until
the supported build dependencies select patched versions themselves.
Joe approved a scoped exception for GHSA-vfj7-8cjw-p6xm (braces <=3.0.3 stack exhaustion). No patched braces version is currently available. Tailwind 3 uses this package only while building CSS from trusted repository paths; the Docker runtime ships compiled CSS, not these npm dependencies. Untrusted brace patterns must not be supplied to the build tools. This is an accepted build-time risk, not a vulnerability fix or a clean audit claim.
scripts/audit_frontend.py accepts only this exact advisory and its transitive
reports in braces/chokidar/micromatch/fast-glob/tailwindcss, requiring every affected
lockfile node to remain development-only. Other moderate/high/critical advisories,
missing audit data and audit failures still fail verification. The exception is
printed on each run. Revisit when upgrading Tailwind or when a patched braces
release is available; remove it once the dependency chain is fixed. Raw npm audit
continues to report the advisory.