Skip to content
autonomous host defense

Defense at
machine speed.

A security daemon for Linux servers. It reads your logs, correlates attacks, bans attackers with expiring kernel rules and audits the box every night. Its AI agent investigates incidents and, from v0.7, acts on its own inside limits you set. Automated attacks meet an automated defender. One static binary. No runtime dependencies.

Get notified at release Open-source release coming soon · Apache-2.0
Try the dashboard
tail -f /var/log/warden.logsimulated output
Simulated output · documentation IP ranges
the dashboard

Your server,
at a glance.

Live activity, incidents, approvals and bans in one view. Approve or reject a proposal, simulate an attack, then switch on "AI acts on its own" to see the agent handle the next one without you.

livemode enforceevents 60.2kblocked 3pending 3health oksimulated

Live · Elevated · 9 incidents today

3 decisions are waiting for you.

Events60.2k5 kinds, 100+ sources
Incidents9worst: high
Blocked now3active bans with a TTL
Waiting for you3proposals to decide

Approve or reject a proposal, or simulate an attack. Turn on "AI acts on its own" to see the agent handle it without you.

Needs you proposals policy sent to you

Block 198.51.100.99 for 1hpending2h ago

Rule http-404-flood · 30 events from 198.51.100.99 within 5m

conf mediumT1595.003
Block 198.51.100.77 for 1hpending2h ago

Rule http-path-probe · probe score 4 within 10m (last match "/.git/")

conf mediumT1595.003
Block 203.0.113.41 for 24hpending1d ago

Rule ssh-password-spray · 41 users, 212 failures in 9m

conf highT1110.003

Blocked now expire on their own

  • 203.0.113.2423h 41mrule:ssh-bruteforce
  • 198.51.100.126d 2hrule:cve-probe-answered
  • 192.0.2.20141mfeed:spamhaus-drop

Where it comes from

  • Frankfurt, Germany1 addr32k
  • São Paulo, Brazilblocked4 addrs8.5k
  • Singapore1 addr7.1k
  • Mumbai, Indiablocked23 addrs3.4k
  • Amsterdam, Netherlands1 addr2.4k

Activity peak 6300 / bucket

Simulated dashboard · runs in your browser · documentation IP ranges · autonomy shown as a v0.7 preview

the rule

The model proposes.
The policy decides.

A tiny root process executes. You choose where the AI may act alone; never-touch lists and hard limits hold whatever it says. The code that reads attacker-controlled logs never runs as root, so even a prompt-injected daemon can only request expiring bans, which the enforcer checks against its own never-touch list.

Policy mode
Throw a curveball

Whatever you pick, the model only ever proposes. Policy, hard limits and the enforcer's never-touch list get the final word.

warden · decision chainsimulated output
incident #118 · source 203.0.113.24 · 212 failures / 9 min · GET /?f=../../etc/passwd
detectssh.spray + web.traversal → incident #118
analystautomated scanner · proposes: ban 203.0.113.24 24h
limitswithin hard limits · single source · expiring
policyauto · rule web-traversal matched
enforcernftables ban 203.0.113.24 · expires in 24h
Enforcer: nftables ban · expires in 24h.
Simulated output · illustrative incident
under the hood

Logs in.
Expiring bans out.

Unprivileged daemonreads attacker-controlled logs · never runs as root
  1. 01logs
  2. 02parse
  3. 03enrich
  4. 04detect
  5. 05correlate
  6. 06incident
analystoptional LLM · writes the verdict from code-computed evidence · never bans
proposals from detectors · analyst · chat · SOAR
Policy engineautoaskdenydry-runfirst match · hard limits no rule can override · you approve "ask"
privilege boundary
root enforcer → nftablesCAP_NET_ADMIN only · no network · checks its own never-touch list · bans expire
Logs are parsed, enriched, run through detection and correlation, and become incidents. An optional LLM analyst explains them. Every proposal goes to a deterministic policy engine, and only a tiny root enforcer, across a privilege boundary, writes expiring nftables bans.
what it does

Detect. Block. Know.
Defend on its own.

01 / DETECT

Detect

SSH brute force, password spraying, logins that succeed after many failures, web scanners, path probes, traversal and injection. Baselines catch low-and-slow distributed attacks. YARA finds webshells and ties them to the requests that touched them. Rules use the Sigma correlation vocabulary, with tests and ATT&CK labels.

~3,000path probes covering 2,500+ CVEs, and a flag when your server answered one
02 / BLOCK

Block

Expiring nftables bans that cover Docker-published ports and survive a reboot. Repeat offenders escalate from 1h to 24h to 7d. Your CrowdSec bouncers can enforce the same bans at the proxy.

03 / KNOW

Know

Spamhaus, abuse.ch, blocklist.de, CINS, ET and Tor exits for free; TAXII, OpenCTI and MISP feeds; network owner and country; CISA KEV and EPSS scores on the CVEs you're exposed to.

04 / AUTONOMOUS AI

Autonomous AI acts on its own from v0.7

Two layers: an LLM agent that investigates incidents, writes a daily briefing and turns false positives into narrower rules, and a decision-model scorer that answers typed questions in milliseconds. From v0.7, the agent acts without waiting for you, under a policy rule you write and only where it was measured accurate on your host.

Secondsfrom AI verdict to kernel ban, no human in the loop
05 / AUDIT

Audit

Nightly read-only checks of sshd, open ports, kernel settings, file permissions and pending security updates, ranked by the exploited CVEs they fix. warden harden applies the fixes with automatic rollback.

06 / INTEGRATE

Integrate

Export incidents to Splunk, Elastic, Sentinel, OTLP or syslog. SOAR and EDR tools connect with scoped, revocable keys.

07 / FLEET

Fleet

A thin hub for a handful of hosts: a ban you approve on one server is offered to the others, through their own policy.

built for AI-automated attacks

Attackers automate.
So does Warden.

MACHINE SPEED

Fast on both sides

Rules ban on the first matching request, in milliseconds. The AI picks up what rules miss and, from v0.7, acts in seconds without waiting for a human.

PATTERNS

Sees the shape of automation

Baselines catch distributed, low-and-slow spraying. CVE probe floods, rotating scanners and known attack infrastructure are flagged from the first request.

INJECTION

Can't be talked out of defending

Attacker text reaches the AI only fenced as data. No tool can unban or allowlist, and hard limits hold whatever a prompt says.

autonomy on your terms

The AI acts where
you let it.

  • Starts in dry-run. It records what it would do until you switch it to enforce.
  • New rule packs and feeds start in monitor mode and are backtested on the last 7 days of events before they can block.
  • The AI acts alone only under a policy rule you write, and only for a model measured accurate on your own host. v0.7
  • Hard limits no rule can lift: one address per incident, bans of 24 hours at most, a separate rate limit, and an automatic stop after you undo three of its decisions. v0.7
  • Chat proposals always wait for you. No model can unban, allowlist or change your configuration.
where it fits

Next to the tools
you know.

fail2banCorrelated incidents, approvals, threat intel, audits and AI triage, not just a regex and a ban.
CrowdSecLocal policy and human approval, nothing leaves the host by default, and it reuses CrowdSec's bouncers.
WazuhOne binary instead of a manager stack, with blocking and AI built in.
FalcoTurns runtime signals into correlated incidents and bans.
Suricata, ZeekComplementary: turns their alerts and flows into correlated incidents and bans. v0.6
WAF, CDNComplementary: host context, audits, and backtested virtual patches for your WAF. v0.7

Only Warden connects attack and exposure on the same host:a probe for a CVE you haven't patched yet, a request for a file YARA just flagged, an exploited CVE still running in an old process.

integrations

Plugs into the stack
you already run.

SOAR playbooks and EDR/SIEM tools connect with scoped, revocable keys. Incidents export to your SIEM in OCSF. Existing CrowdSec bouncers for nginx, Traefik, Caddy, HAProxy or Cloudflare can enforce Warden's bans at the proxy.

SOAREDR / SIEMSplunkElasticSentinelOpenTelemetryMISPOpenCTInginxTraefikCaddyHAProxyCloudflare
roadmap

What's next,
labelled.

v0.6 PlannedCoverage and network visibility. Reads your Suricata or Zeek sensor to see every port: port scans, sweeps, floods and IDS alerts become incidents, and spoofed traffic can't trigger a ban. IPv6 prefix bans, container posture, osquery and Falco.
v0.7 PlannedAutonomous AI, runtime exploits, virtual patching. The AI agent acts on its own under your policy. Exploit detection for nginx, PHP and the JVM. CVE virtual patches, backtested on your traffic, exported to your Coraza WAF or as Suricata rules.
v0.8 PlannedSmarter decisions. Confidence-based actions: block, short ban with a notification, or watch.
v0.9 PlannedKernel prevention. eBPF telemetry, exec blocking, memory protection, tamper resistance, and banning the real client behind a CDN.
v1.0 PlannedFleet. Signed rule feeds, cross-host consensus, outside-in exposure scans.
status

v0.5.0, no spin.

PlatformsLinux amd64 and arm64 with systemd and nftables; packages for Debian, Ubuntu and the RHEL family
TestingEnd to end in containers and in a VM lab under systemd (Debian 12, Ubuntu 22.04 and 24.04) — not yet proven on busy internet-facing servers
AIToday every AI proposal waits for your approval. Acting on its own, under your policy, arrives in v0.7
LicenseApache-2.0 · open-source release coming soon
Built forSmall teams running Linux servers without a security operations center: devops at SaaS companies, agencies running client servers, VPS operators
open and enterprise

Free on every server.
Fleets are enterprise.

The daemon, enforcer, policy engine, rules, AI agent and dashboard are the open-source core. Planned for teams: a fleet console with central policy, premium rule packs and intel feed, a hosted LLM analyst, enterprise connectors, and LTS builds with support.

Planned
release

Open source, soon.

Warden's source is being prepared for public release under Apache-2.0.