Trust
Safe by default. Auditable always.
Vigilhold watches process, network, file, and SSH signals on Linux servers, including Docker hosts. It correlates what it sees and, on hosts where you configure it, can act — and when it does, every step is traceable, reversible where possible, and never hidden.
How Vigilhold Works
Vigilhold works in five steps: it watches, investigates, decides, acts where you allow it, and reflects on the outcome.
- 1 Watch Collect live telemetry from every agent — metrics, processes, network, logs
- 2 Investigate Open a case, correlate signals, map to MITRE ATT&CK techniques
- 3 Decide Score confidence, select strategy, check approval requirements
- 4 Execute Act — ban IP, kill process, quarantine file, isolate network
- 5 Reflect Evaluate outcome, update confidence, surface learnings
These steps run on Cortex, Vigilhold's decision engine: written rules and scoring, not a trained AI model. Where it is switched on, an AI model (Anthropic's Claude) writes the longer explanations; it never decides what to do. Every decision includes a confidence score, and your autonomy level decides what that score may do: on Autopilot a high-confidence decision acts and tells you, and anything less waits for a person. On Observe and Assist the limits below apply whatever the score.
Autonomy Levels
You choose how much Vigilhold acts on its own. New organisations start on Assist; switch to Observe at any time, or up to the highest level your plan includes, and every change is in the audit log.
ObserveVigilhold detects and logs. Nothing executes automatically. Full visibility, no autonomous action.
Assist DefaultTemporary IP bans and rate limits run automatically, along with steps that only read or notify. Everything else that changes a host (stop a process, isolate, lock down) waits for your approval.
AutopilotBusiness and Enterprise. High-confidence threats are contained without waiting for you, within your hourly action limit and the actions you keep for approval. Bans, quarantines, isolation and lockdowns can be undone.
SovereignAn extra level name on Enterprise. Today it behaves exactly like Autopilot, with the same limits and the same undo.
Isolation and lockdown: the rule and its exceptions
- On Observe and Assist, Vigilhold's detections never isolate a server on their own: they raise the alert and wait for a person.
- On every level, isolation or lockdown that comes out of a Cortex investigation plan waits for a person.
- On Autopilot, the most serious confirmed detections may isolate a server on their own.
- On Autopilot, if many of your servers go silent at once, the ones still connected may lock down for a while. This is a deliberate exception: it protects them while nobody can reach the fleet.
- A lockdown playbook or threat-score trigger an admin turned on runs by itself on Assist and Autopilot. Lockdown playbooks start off.
Human Overrides
Overrides are also feedback. Each correction counts in future decisions on that server and that kind of threat.
What overrides record
- Who made the override (username + role)
- When they made it (timestamp)
- Why they made it (free-text justification)
- What signal triggered the original decision
- The confidence score at decision time
- Whether the action was rolled back
What overrides do
- Block or approve a pending action
- Lower Vigilhold's confidence on similar events on that server
- Mark an action as false positive (updates reflect log)
- Add to the suppress list to avoid repeat noise
- Trigger rollback if the action has already executed
- No Override another user's override without audit trail
Two-person authentication is available on Enterprise plans. When two-person approval is on, kills, bans and lockdowns an operator starts need a second operator; automatic responses follow your autonomy level.
Rollback
Every action in the registry is tagged reversible or irreversible. An irreversible action, such as stopping a process, runs on its own only on Autopilot and only at high confidence, adjustable per threat type, on hosts whose limits allow responses. Below that it waits for a person, and it never runs just because nobody answered. On Assist and Observe it waits for your approval, unless an admin chose the aggressive response setting, which also stops likely reverse shells.
IP ban TTL-based auto-expiry or manual unban Reversible Process kill (SIGKILL) Process must be restarted manually Irreversible Network isolation Release isolation command sent to agent Reversible File quarantine File returned to original path with reason logged Reversible Lockdown Lockdown lifted via control plane command Reversible SSH key revocation Key must be re-added manually IrreversibleRollback requests are logged with the same chain of custody as the original action.
What Runs on Your Server
Installing needs root. Afterwards Vigilhold keeps two services running on a host. The agent watches and reports, as an unprivileged user. The warden, a small separate service, runs as root and does a fixed set of tasks on the agent's behalf under rules no command can override, so even a compromised agent can't get past them. Short-lived root helpers install signed updates, apply hardening you approved and keep a read-only copy of root-only files for the agent. The installer also adds malware scanners (ClamAV, and rkhunter where available); ClamAV scans run as root from a timer, every 6 hours at the lowest priority. The install guide lists everything the installer adds.
The agent: its own user, not root
- Runs as the
sentinel-agentuser with five Linux capabilities: two to manage the firewall and raw network sockets, and one each to send signals to processes, read other processes' details, and read the kernel log. It cannot gain more (NoNewPrivileges), and theptracesystem call itself is blocked. - Cannot change its own code. Its files are owned by root and made immutable, and that is reapplied before every start.
- Fenced in by systemd: read-only system directories, read-only home directories, a system-call filter, no new namespaces, and hard caps of 10% CPU, 1 GB of memory and 64 tasks.
- Today the agent also keeps its own firewall and kill permissions, as a fallback for hosts where the warden isn't running. We plan to remove them.
The warden: root, within fixed limits
- A small Rust program that does a fixed set of actions — address blocks, process stops, file quarantine, hardening you approved, account locks and updates — under rules no command can override.
- Only answers the agent's own service, checking the caller's user and systemd unit on every request, and root, which can already do all of this.
- Never blocks Vigilhold's own servers or local networks, or stops the services that keep the host reachable, and checks a process is still the one it was asked about.
- Rate-limited, and every request is written to a root-only log.
Updates
- Every release is signed with an Ed25519 key held only on Vigilhold's build machine, never on the servers that hand out updates. A root updater checks that signature against a key built into it and refuses anything unsigned or older than what is installed.
- If the new agent crashes or doesn't finish starting within two minutes, the updater puts the previous version back on its own.
- New releases reach early hosts first. Other hosts follow within a day after an early host runs it cleanly; urgent security fixes move faster.
Commands and evidence
- Every command to a host is signed with that host's secret and numbered, so the agent rejects a forged or replayed one. Changing where a host connects needs a second signature from Vigilhold's signing key.
- Alerts and coverage reports are signed with Vigilhold's evidence key, published at
/api/evidence/keys. Anyone can check them offline with our verifier. sudo vigilhold-uninstallremoves the agent, the warden, the updater and every service they installed, and lifts the blocks the agent placed. It lists what it leaves and how to remove it: packages that may be yours, and hardening settings.
What's Live Now
Vigilhold is honest about what ships. These are the features in production today.
Feature Status Real-time agent telemetry (CPU, memory, disk, processes, network) LIVE Autonomous brute-force detection and IP ban LIVE Cryptominer detection, with a one-click process stop LIVE SSH configuration baseline and drift detection LIVE File integrity monitoring (FIM) LIVE Investigation pipeline with confidence scoring LIVE Human override queue with approval workflow LIVE Two-person approval for operator-started kills, bans and lockdowns (Enterprise) LIVE Simulation mode — see what would run before anything does LIVE Compliance evidence packs (SOC 2, CIS, PCI-DSS) LIVE Hash-chained audit log with chain verification LIVE Time-to-detect and time-to-contain metrics (MTTD, MTTC, MTTR) through the API LIVE Decision trace — full signal-to-outcome provenance LIVE Integration dispatch (Slack, PagerDuty, Jira, Splunk, Webhook) LIVE Hardware-backed key vault (HSM) ROADMAPPerformance
The Cortex step that turns a context snapshot into a threat verdict runs in memory, with no database queries or network calls, so scoring a threat does not hold up a response.
Data & Audits
What data Vigilhold collects
- System metrics (CPU, memory, disk, I/O) from each agent
- Running process list (name, PID, CPU/mem, command line). Command lines are encrypted with a per-organization key
- Network connections: listening ports, and active connections with each outbound destination as a fingerprint made on your server with a key that never leaves it. Your own servers and addresses on public threat lists are shown as addresses, and so are inbound attackers and a reverse shell's destination, because you have to act on them.
- Failed and successful login events from system logs, with username and source address
- New lines from the system's authentication, system and kernel logs as they are written, so the evidence survives if an attacker wipes the files, and system error messages
- The text of a short list of system configuration files, when the agent connects and as line changes afterwards: SSH, sudo, cron, /etc/hosts, /etc/passwd and nginx site settings, with values that look secret blanked
- Installed packages, services and containers
- Hashes, paths and change events for monitored files
- From outside: Vigilhold's servers check, with read-only requests to your server's public address, whether risky ports or secret files such as /.env are reachable. File contents are never kept, and you can turn this off for any server.
What Vigilhold collects on request
- A forensic collection, when you start it, and automatically when Cortex confirms an intrusion or before it contains a server: recent authentication log lines, SSH and cron configuration, loaded kernel modules, files open from temporary or deleted paths, and the last commands in shell histories, with typed passwords, tokens and keys blanked out on the server first
- Log tails, error collection and repair output when you ask for them
- Terminal sessions you open from the dashboard, which are recorded
What Vigilhold does not collect
- No Other file contents, such as your application code, data or home directories: Vigilhold sees those only as hashes, paths and change events
- No Keystrokes outside the terminal sessions you open from the dashboard
- No Application-layer payloads or HTTP request bodies
- No Passwords, keys and tokens on purpose. Blanking works by pattern, so a secret in an unusual format, or one passed as a command-line argument, can still appear in a log line or process list
- No Environment variable values (an unexpected preloaded library, from LD_PRELOAD or /etc/ld.so.preload, is named in its alert)
Audit log guarantees
Every admin action — including every automated decision, approval, override, and rollback — is written to a tamper-evident audit log: a SHA-256 hash chain. Each entry includes a chain_hash and prev_hash that link entries cryptographically; each entry's hash covers what happened, when, who did it, on which server and from which address. Tamper detection is built in: compliance evidence bundles include chain verification status so auditors can confirm log integrity independently.
Data residency
Every row is tagged with its organization and every query is scoped to it. Warnings between your own servers stay inside your organization. Sharing with other organizations is off unless you opt in, and then only attacker details leave: no org_id, server identifiers, or raw payloads. Single-tenant deployments: contact us.
Who Can Do What
Every Vigilhold action is gated by three independent checks that all must pass: (1) org role, (2) plan entitlement, and (3) server ownership. Role defaults are shown below. Owners and admins can grant or revoke individual permissions per member on top of these defaults.
OwnerFirst user in the org. All permissions including billing and plan management. The last owner cannot be removed or demoted.
AdminDelegated admin. All operational permissions except billing. Can invite, demote, or remove other members.
AnalystSOC / security staff. Can view audit log and manage alerts. Cannot kill processes, ban IPs, or take destructive actions.
ViewerRead-only. Can see fleet health, alerts, and metrics. Cannot acknowledge, approve, or take any action.
Permission matrix — role defaults
| Action / permission | Owner | Admin | Analyst | Viewer | Plan required |
|---|---|---|---|---|---|
| Read-only | |||||
| View fleet, metrics, processes | All | ||||
| View alert list | All | ||||
| View pending approvals | All | ||||
| View audit log (filtered) Analyst sees filtered view — source IPs stripped | — | All | |||
| View full audit log + source IPs | — | — | All | ||
| Alert & response actions | |||||
| Acknowledge / manage alerts | — | All | |||
| Approve / reject queued actions | — | — | All | ||
| Kill processes | — | — | All | ||
| Ban / unban IPs | — | — | All | ||
| Enable / disable lockdown | — | — | All | ||
| Configuration | |||||
| Manage alert rules & playbooks | — | — | Developer+ | ||
| Manage integrations & webhooks | — | — | Business+ | ||
| Configure SSO | — | — | — | Enterprise | |
| Configure zero-trust policies | — | — | Enterprise | ||
| Team & access management | |||||
| Invite & remove team members | — | — | All | ||
| Change member roles | — | — | All | ||
| Grant per-member permission overrides Admins only within their own permissions | — | — | All | ||
| Manage billing & plan | — | — | — | All | |
| Security & secrets | |||||
| Create / revoke API keys Owner only | — | — | — | All | |
| Read vault entries | — | — | Developer+ | ||
| Write vault entries | — | — | Developer+ | ||
| Generate emergency access token | — | — | Business+ | ||
| Generate compliance reports | — | — | Business+ | ||
| Export the audit log (CSV or JSON) API key requires read:audit scope | — | — | All plans | ||
| Verify audit chain integrity Owner only. Requires full auth — emergency-access tokens rejected | — | — | — | All |
API key scope gates
Scope What it grants Minimum plan<area>:read Read one area: servers, metrics, processes, alerts, audit, fleet, compliance, webhooks, events, stream All <area>:write Change one area: alerts, webhooks, events; actions:write for ban, scan, lockdown and Cortex approvals All read:* / write:* Every area for that operation — still only the public API All audit:read Required for the audit log and its export All plans (any) No key can reach anything outside the public API: not the vault, terminals, users, billing or settings All System admin — a Vigilhold platform operator bypasses all org-role and plan gates. Every change staff make in a customer's organization is logged.
Per-member overrides — org owners can grant or revoke any individual permission for any member. Overrides stack on top of role defaults. Only owners can elevate a member beyond admin-level defaults.
Permission cache — resolved permissions are cached per member for 60 seconds. Role changes take effect within one minute without requiring re-login.
Emergency access tokens — single-use tokens with a 24-hour window. They carry emergencyAccess: true and are rejected by requireFullAuth — they cannot approve actions or mutate destructive resources.
Enterprise Procurement
This section answers the questions a security team, procurement team, or compliance officer will ask before approving Vigilhold for production use.
Architecture summary
Vigilhold has three components: (1) a lightweight Node.js agent running on each monitored server, (2) a Node.js backend (the Vigilhold dashboard and API), and (3) Cortex, the decision engine (written rules and scoring, not a trained model), which runs on the backend. The agent connects outbound to the backend over authenticated WebSocket (WSS). No inbound ports are required on monitored servers.
Detection runs on the agent. If it loses contact with the backend it keeps defending on its own, within the policy it last received: it bans SSH brute force and applies hardening you approved. Decisions that need the whole picture (Cortex plans, lockdowns, approvals) are made by the backend.
Data handling summary
- TLS on an authenticated, mutually-verified connection; every command to the agent is HMAC-signed and protected commands carry a second signature
- Every row is tagged with its organization and every query is scoped to it
- Credentials in the vault are encrypted with AES-256-GCM, per-org key isolation
- Audit logs are SHA-256 hash-chained — tampering produces a detectable chain break
- Sharing threat data with other organizations is off unless you opt in, and then only attacker details leave: no org_id, server IDs, or raw payloads
- Hosted on Cloudflare and on servers in the EU and US
- Single-tenant deployments: contact us.
Support and SLA story
- Every plan: email support, first reply within one business day
- URGENT in the subject: first reply within 4 hours, any day of the week
- Uptime: a 99.5% monthly target for the dashboard and agent connections, a target, not a guarantee; live uptime and every outage are on the status page, and we publish what went wrong after significant outages
Rollout and onboarding
Agent install is a single curl command per server. It can be scripted via Ansible, Chef, Puppet, or any provisioning tool that can run a shell command as root. Mass deployment to hundreds of servers is supported via bulk agent token creation through the API.
Vigilhold does not require a maintenance window for installation. The agent starts reporting within seconds of install. Existing SSH sessions and running processes are not interrupted.
Known limitations (honest)
- No Windows agents are not available yet; Vigilhold runs on Ubuntu, Debian, RHEL-compatible, Amazon Linux 2023, Arch and Alpine servers, including Docker hosts
- No macOS agents are not available
- No Kubernetes (DaemonSet) deployment is not available yet
- No Vigilhold does not do SAST, DAST, or code scanning
- No SCIM provisioning requires Enterprise plan
Support response times
Email support@vigilhold.com. You get a first reply within one business day (Monday to Friday, US Mountain Time). If you think a server is under attack, or Vigilhold did something you didn't expect, put URGENT in the subject: urgent messages get a first reply within 4 hours, any day of the week.
Questions for your security review
Security questionnaires: contact security@vigilhold.com.
Procurement resources
Permissions Matrix →
Who can do what — full role and permission breakdown
FAQ →
Common buyer questions answered directly
Release State →
What is live, beta, experimental, or roadmap
Vendor SLA →
Uptime target, support times and incident process — request via email
Rollback Playbook →
How to roll back any Vigilhold action or deployment — request via email
Standards & Safeguards
- SOC 2
- Controls in place; independent audit not yet started
- CIS Benchmarks
- Built-in control checks
- PCI-DSS
- Control mapping
- ISO 27001
- Alignment underway
Data residency & sovereignty
Hosted service: each organization’s data is kept separate and is never shared with another organization without explicit opt-in and anonymization. Single-tenant deployments: contact us.
Signed updates
Agent releases are signed, and the agent refuses an update that fails the check. Every command to a host is HMAC-signed and rejected without a valid signature. An administrator can rotate a host’s command key at any time.
No hidden automation
Every automatic containment action appears in the response log and the audit log. Nothing executes silently.
Tamper-evident audit log
A SHA-256 hash chain: every entry links to the previous one, so an edit or deletion breaks the chain, and breaks are surfaced in compliance reports.
Explicit approval paths
What waits for a person depends on your autonomy level. On Observe and Assist, destructive steps wait, with the exceptions listed under Autonomy Levels. On Autopilot, high-confidence steps may act and the rest wait. A step waiting for approval never runs because nobody answered.
Simulation mode
Turn on simulation to see exactly what Vigilhold would do, step by step, before any real action runs. Safe for onboarding.
Visible uncertainty
Vigilhold shows confidence scores, not just verdicts. When confidence is low, it says so and queues for review rather than acting.
Security questionnaires: contact us
Contact security teamQuestions about how Vigilhold handles your data?
We reply to security questions within one business day, and to urgent ones within 4 hours.
Contact security team © 2026 Vigilhold. All rights reserved. Permissions matrixFAQHelp & supportRelease stateChangelogMedia kitSign in