Episode handles device credentials and potentially sensitive evidence. Please do not publish credentials, private camera addresses, raw evidence, or security vulnerability details in a public issue.
During pre-1.0 development, only the most recent tagged prerelease receives security fixes. The project is not yet suitable for direct Internet exposure and does not provide authentication.
Use GitHub's private vulnerability reporting for this repository under Security → Advisories → Report a vulnerability. If that option is not available, open a public issue containing no technical details and ask the maintainer to establish private contact.
Please include the affected Episode version, deployment environment, impact, and minimal reproduction steps. Remove all passwords, tokens, private addresses, and captured evidence.
General bugs and hardening suggestions that do not expose a vulnerability can be reported through the public issue tracker.
Treat Episode, its Event and Alarm Server endpoints, FTP service, camera network, and mounted plugin directory as one trusted local security boundary. Third-party plugins execute as trusted code inside the Episode process and are not sandboxed. Review them before installation and keep the plugin mount read-only.
The unauthenticated API also exposes operational configuration, including the active Capture profile. A client with network access can select a restrictive or empty profile and prevent later active Events from starting capture. Limit UI/API access with the host firewall or a trusted reverse proxy until Episode provides authentication and authorization.
Filtering event classes reduces capture coverage, so it is treated as a
security-relevant configuration change rather than a noise preference. Every
event class is selectable at both the Capture Profile and the Device level: a
profile may suppress security (tamper_detection, tampering_detection,
video_loss, battery_low), a classified detection or rule including a wired
alarm input (digital_input, which both the Hikvision alarm report and an
ONVIF digital-input topic reach), an access record, or unknown. Suppressing
detection therefore also silences a wired trigger, so it is not a safe "drop
the object detections only" choice. That consistency is deliberate — one selector has to
mean one thing, whatever it names and whoever it is applied to — so protection
comes from the operator seeing the consequence, not from a class the validation
refuses. Selecting security for a camera is marked in the Device list while it
is active, and any filtering shows in the profile and Device summaries.
Filtering never removes evidence. The Raw Artifact, Receipt, and canonical Event are persisted first, and a filtered Event keeps its reason, suppressed type, class, deciding level, and profile. An operator who disabled filtering later cannot recover capture that was never started, so the active profile and any per-Device selectors are part of the deployment's security posture and should be reviewed with it.
The plugin-side ignore_events setting (for example Hikvision ISAPI) is a
different lever: it stops a vendor message from being interpreted, so no
canonical Event is produced at all. Core filtering stops capture after an
Event exists. In the current ISAPI implementation, repeated identical ignored
states are also suppressed before storage; those repeats have no Raw Artifact
or Receipt. Prefer core filtering; use ignore_events only when a device
emits messages the core should never turn into Events.
Treat an Episode-start webhook URL as a credential: Discord and similar URLs contain bearer-like secret tokens. Configure it through System → Notifications, do not publish it in diagnostics or issues, and rotate it if it is exposed. Episode stores the URL in its local SQLite settings but does not return or log it; the UI displays only a fixed mask and whether a destination exists. Prefer HTTPS; plain HTTP is appropriate only inside a trusted network where interception is acceptable. The notification payload contains operational Episode, Area, Device, Event, and timestamp information, but no raw payload, Evidence bytes, camera credentials, or absolute filesystem paths. The optional global External Episode URL is not secret and is returned by the installation settings API; it is operator-supplied rather than inferred from an untrusted request host and is validated without credentials, queries, or fragments.
Recording diagnostics are credential-free for the same reason. FFmpeg prints
the input stream URL verbatim when it cannot open it, and that URL contains the
camera username and password, so the recorder strips the user:password part
before it keeps any of FFmpeg's output. Only the redacted tail reaches a log
line, a recording's reported reason, or System → Recordings; host and port
survive because they identify which camera failed. Raw FFmpeg output is never
persisted.
SHA-256 checksums detect accidental or later byte changes; they are not a signature, trusted timestamp, or legal chain of custody. Signed manifests and external timestamping remain future hardening work.