Security
We don’t charge extra for trust.
Every ModelReins deployment — free, pro, team — ships with the full security stack. No feature gates. No “upgrade to see who’s attacking you.” Security observability is how an orchestrator proves it’s trustworthy.
Layered defense
Section titled “Layered defense”| Layer | What it does |
|---|---|
| Edge protections | Managed challenges on unauthenticated API probes. Endpoint enumeration is discouraged before requests reach your origin. |
| Host-level brute force protection | Repeated failed logins from the same source get a temporary ban. Tuned so a fat-fingered password is forgiven; scripted attempts are not. |
| Uniform 401 responses | Unauthenticated requests all return the same thing. Real endpoints and fake endpoints look identical from the outside. |
| Consolidated public surface | Display data comes through a single authenticated endpoint. Client code never reveals a map of internal paths. |
| Circuit breakers | External dependencies open a circuit on repeated failures, preventing cascade and runaway cost. |
| Audit trail | Every action is logged with HMAC signatures. Tamper-evident. Queryable by actor, action, or time range. |
| Vault-backed credential isolation | Tenant secrets live in a dedicated vault, not the app database. Resolved at runtime with short-TTL cache. |
| SSRF protection | Connectors refuse internal network ranges. Dangerous headers are stripped before any outbound request. |
Wall of Shame
Section titled “Wall of Shame”Every IP that tries to break into your fleet gets logged with:
- IP address
- Country (GeoIP)
- ISP / ASN
- Which host they hit
- Which service
- How many failed attempts
- When they were banned
Sentence: death. Bans are applied in real time. The Wall of Shame is a live billboard page — visible on the Companion screensaver and the web dashboard.
Typos vs. attacks
Section titled “Typos vs. attacks”Mistyping your password once doesn’t get you on the Wall. The threshold is tuned to distinguish fat fingers from brute force:
- A couple of failures from a legitimate user — nothing happens. We’ve all been there.
- Repeated failures from the same source over a short window — that’s not a typo. Banned.
- Distributed attempts (different IPs, same target) — caught at the edge before reaching the host layer.
If you do manage to lock yourself out, bans are temporary and expire automatically. Your admin can also unban instantly from the dashboard. You won’t end up on the Wall of Shame — that’s reserved for sources that have no legitimate reason to be trying.
The Wall of Shame is not a premium feature. It ships with every deployment.
What’s included at every tier
Section titled “What’s included at every tier”- Full audit trail with HMAC signatures
- Wall of Shame with GeoIP
- Brute force protection and reporting
- Circuit breakers for all external services
- Edge anti-enumeration rules
- SSRF protection on connectors
- Vault-backed credential isolation
- Alerts on security events via your preferred channel
Killswitch
Section titled “Killswitch”Graduated abort levels for stopping work in progress:
| Level | Scope | Where |
|---|---|---|
| 0 | No-op | — |
| 1 | Cancel current prompt | Saddle — click the stop button mid-stream |
| 2 | Abort thread | Saddle — red button in the command strip kills the active thread and all pending dispatches for that thread |
| 3 | Pause worker | Dashboard — pause button on a worker card stops it from picking up new jobs |
| 4 | KILL ALL | Dashboard — tenant-wide abort. Cancels every in-flight job, pauses every worker, drains the queue. |
Level 2 (thread-scoped abort) is the most common. It stops the current chain without affecting other threads or workers. Level 4 is the emergency brake — use it when something is burning tokens faster than you can think.
After a KILL ALL, workers stay paused until you manually resume them from the dashboard.
What’s NOT included
Section titled “What’s NOT included”Charging you more to be safe.