Skip to content

ReD-Pot 1.0.0

Choose a tag to compare

@github-actions github-actions released this 13 Nov 19:09
· 1 commit to main since this release

First release of ReD-Pot Community Edition (MIT), built from v0.4.0. It keeps every decoy role, the baits and
honeytokens, scoring, the dashboard and remote sensors, and sets the Community Edition limits: at most 3 decoys, one
decoy per network interface on that interface's own IP, a fixed persona and bait set, no outbound alerting and no
offline replay. See "Community Edition limits" in the README.

Upgrading: fresh installs only. Uninstall v0.3.x / v0.4.x (sudo ./install.sh uninstall), then install 1.0.0
(new persona, baits and admin token; remote sensors join again). Commands that change an existing install refuse a
configuration made by an older version.

Added

  • Export: Incidents → Export JSON downloads incidents and alert records for a date range (whole UTC days, To
    included, at most 366 days) as one JSON file with a "complete": true marker. Admin sessions only; one export at a
    time; larger than 200 MB is refused with the counts instead of being cut off.
    The same from the command line: sudo ./install.sh export --from D --to D [-o FILE] [--force] (file mode 0600, never follows symlinks, --force replaces only a regular file), and inside the
    console container rpotctl export (JSON on stdout) and rpotctl slots (decoy slot counts).
  • Decoy limit (3 decoys: local decoys + remote sensors + pending joins), enforced by the installer, rpotctl,
    enrollment, the database and the sensor API. The dashboard shows "n of 3 decoys", disables Add a sensor when full
    and explains why; pending joins can be cancelled, which frees the slot at once. Revoking a remote sensor frees its
    slot; a local decoy frees its slot only with install.sh remove-decoy. A full manager answers a valid join with
    "decoy limit reached" and the token stays unused.
  • Edition label in the dashboard header; edition in /api/summary, slot counts in /api/fleet.
  • sensor.gap note: after an outage a remote sensor reports how many events it could not deliver (count and first/last
    timestamp only), shown on that decoy's incident.
  • Alert records are stored in the database (alert_records) as the source for Export, whether or not alerts.json
    is enabled. They follow the event retention.
  • tests/lab_scenarios.sh setup|attack|all|down builds a 3-decoy bridge lab (FS01, DC01, WEB01) and runs the simulated
    workstations; DRY_RUN=1 prints the commands.
  • Release pipeline (GitHub Actions): CI on every push and pull request (tests on Python 3.12/3.13, CE gate, shell
    syntax, dependency audit). A version tag builds the image for linux/amd64 and linux/arm64, publishes
    ghcr.io/redsidesecurity/red-pot:<version> with provenance and SBOM, checks the pushed image's source-hash label
    against the installer's src_hash(), then creates the GitHub Release from this changelog.

Changed

  • SSH decoy: an open interactive shell is no longer closed when no further channel is opened within 60 s.
  • Connection floods (pentest F-02): one decoy-wide connection budget for every TCP service, SMB and HTTP included
    (250 concurrent, 20 per source IP, thread ceiling 320 under the container's 384); SMB sessions without a named logon
    idle out after 60 s. A listener no longer dies when the decoy runs out of threads: that client is dropped and the
    listener keeps serving. Sources that hit a cap raise a level-3 conn.flood event (once a minute per source).
  • SMB named-pipe flood (pentest N-3): SMB named pipes (\srvsvc, \wkssvc) no longer start a thread each; RPC
    requests are answered in the SMB session's own thread. Every open pipe is charged to the decoy's connection budget:
    at most 8 per SMB session, 16 per source IP, 128 per decoy, and none once the decoy is within 32 threads of its
    thread ceiling (a pipe is refused before a new client is). A refused open gets STATUS_INSUFFICIENT_RESOURCES; the
    session keeps working. Idle pipes close after 60 s (null session) or 300 s (named logon), and all of a session's
    pipes close with the session. The SSH decoy allows 10 channels per connection (OpenSSH MaxSessions), charged the
    same way.
  • Resource-exhaustion detection: a source that hits its own pipe cap raises smb.pipe_flood (level 4); its own
    SSH channel cap, or holding all its connection slots while the decoy is full, raises resource.exhaustion (level 4).
    When the decoy-wide thread or connection cap refuses a client, the manager opens or raises a level-4 decoy incident
    "Decoy NAME saturated - possible DoS". That incident names the top holders (source IPs and their counts), not the
    refused client, and always writes an alert record when it opens or rises (otherwise at most one per decoy per 5 min).
    Flood events now carry the total refused since the flood started (refused), the count since the last report
    (new) and duration_s. The first refusal is reported at once and a summary follows within 60 s.
  • Decoy service health: every 30 s each decoy checks its listeners; a dead or unbound one is restarted and reported
    (sensor.service_down / sensor.service_up). The manager opens a level-4 "service down - possible DoS" incident on
    that decoy and /api/sensors shows services_down and per-service health; the dashboard marks a down service in red.
  • Persona: the installer asks only the AD/DNS domain and the industry. Company name, NetBIOS name and mail domain
    are derived from the domain and shown on the review screen. Users come from a larger built-in name pool (70 first
    names × 72 last names), 14 picked per install by a random seed, which also picks the username format. Internationalised
    (xn--) domain labels are refused with a hint. --domain is required with --yes (no default domain).
    The NetBIOS name is the first domain label without hyphens (corp.ararat-bank.am → ARARATBANK); when that is
    longer than 15 characters, its first word (northwind-traders-intl.com → NORTHWIND). Before: ARARAT-BANK,
    or a 15-character cut such as REDSIDE-SECURIT (pentest F-04).
  • Fewer honeypot tells (pentest F-04, F-05):
    • Bait files are dated per file at office-hours times with their own minutes and seconds (before, every file carried
      the install time's seconds). .docx baits have Word-like metadata: author and editor from the staff list, created
      and modified dates matching the file date, app.xml, and ZIP entries dated like Word's (no install time).
    • The DC's LDAP directory looks like a real Server 2019 domain: about 125 objects, including the default containers,
      CN=Builtin with the builtin groups, the default domain groups, the Default Domain (Controllers) Policy GPOs,
      a company OU tree with department OUs and groups, and one workstation per employee (also known to the KDC).
      Every memberOf points to an existing object, member matches it, and the primary group is not listed.
    • A null session can no longer list shares (Server 2019 default). Logged-on bait users still can.
    • The NetBIOS node-status reply carries the decoy interface's real MAC address instead of a made-up one.
  • One decoy per interface: each decoy answers on its interface's own IP; the address question is gone. The
    installer offers each interface once and allows at most min(3, usable interfaces) decoys. A 1-NIC server holds one
    local decoy, which shares the management IP (host SSH moves to a random port). --no-ssh-move with a decoy on the
    management interface is an error. Interfaces with unusual names are listed as Not usable.
  • add-decoy offers only interfaces without a decoy, checks the decoy count with the running manager first, and needs
    the manager to be running.
  • Remote sensor install: no placement or address question; the decoy answers on the IP of the chosen decoy interface.
    On the manager-link interface host SSH moves (and the decoy follows DHCP there); on another interface the decoy binds
    that IP and host SSH is pinned to the manager-link IP. The decoy-interface firewall policy applies on sensors too.
  • Manager unreachable: a sensor retries the batch in flight in memory for up to 60 s; after that, events go only to
    a write-only local log undelivered-<NAME>.jsonl on the sensor (mode 0600, RPOT_UNDELIVERED_MAX_MB, default 50) and
    are never sent later.
  • Sensor timestamps older than 15 minutes (was 7 days) or more than 5 minutes in the future are replaced by the
    manager's time; the original is kept as data.sensor_ts, and the manager log warns once per sensor about clock skew.
  • rpot.json format version 2. alerts keeps only min_level (the alerts.json threshold).
  • rpotctl init takes --domain (required) and --profile; --net-mode is host (production,
    --sensor NAME=role@IP%IFACE) or bridge (isolated lab). At most 3 --sensor entries.
  • The version (1.0.0) has a single source, rpot/__init__.py; the installer, rpotctl and the image tag read it.
  • DC DNS fingerprint: the domain controller decoy's DNS now answers like a default Windows Server DNS: CHAOS
    queries (version.bind) get NOTIMP with no version, other answers have RA set (header bit only; the decoy still
    never recurses), and non-QUERY opcodes (DNS STATUS) get a 12-byte NOTIMP. nmap -sV reported it as
    "NLnet Labs NSD"; it now reports "Simple DNS Plus" (OS: Windows), as it does for most real domain controllers.
  • DNS zone apex: the DC and Linux DNS decoys answered NXDOMAIN for the domain itself (SOA, NS, A) and for
    _msdcs / _tcp while names below them resolved, and negative answers had no SOA. The DC now serves the domain
    zone and an _msdcs zone with Windows defaults (SOA primary = the DC, hostmaster.<domain>, 900/600/86400/3600),
    NS and apex A = the DC, the Net Logon records (site, pdc/gc, DC and domain GUID records, DomainDnsZones /
    ForestDnsZones). The Linux DNS role gets a BIND-style SOA and NS. Empty names such as _tcp answer NOERROR with
    no data, NXDOMAIN/NODATA carry the SOA, zone transfers stay refused, and nothing is resolved outside the zone.
  • RDP fingerprint: the RDP decoy now looks like Windows Server 2019 with Network Level Authentication required.
    TLS is capped at 1.2 (no TLS 1.3) with the Schannel cipher order and 2048-bit DHE. The certificate is a per-install
    self-signed one like the one Remote Desktop creates for itself: CN = host FQDN only, RSA 2048 / SHA-256, a 16-byte
    serial, 6-month validity (re-issued when it expires), EKU Server Authentication and Key Usage Key Encipherment +
    Data Encipherment, and no other extensions. A client that offers only Standard RDP Security or plain TLS gets an
    RDP_NEG_FAILURE with HYBRID_REQUIRED_BY_SERVER instead of a silent close. RDSTLS and CredSSP with Early User
    Authorization are accepted, and a bare TLS handshake on the RDP port shows the certificate, so nmap -sC reports
    ssl-cert on 3389 like on a Windows host.
  • LDAPS and HTTPS fingerprint: the domain controller's LDAPS (636) and the new Global Catalog over SSL port
    (3269) present a certificate issued by a per-install fake enterprise CA named like the Active Directory Certificate
    Services default (<domain>-<DC name>-CA, its key stays in the sensor data directory, mode 0600). The certificate
    follows the "Domain Controller" template: CN and DNS name = the DC FQDN plus the DC object GUID, EKU Client +
    Server Authentication, template name DomainController, LDAP CRL and AIA locations, the DC account SID, an AD CS
    style 19-byte serial and a 1-year validity from a backdated issue date. TLS follows Windows Server 2019 (1.0 to 1.2,
    no 1.3, Schannel cipher order). The Linux web decoy's HTTPS now looks like Apache's default SSL site on Ubuntu
    22.04: the "snakeoil" self-signed certificate, TLS 1.2 and 1.3 only, HIGH:!aNULL with client cipher order.
  • Web decoy errors look like Apache (pentest N-4): unusual or malformed requests no longer get Python
    http.server pages. Error responses use Apache 2.4.52's error documents, reason phrases and header order
    (Date, Server, ..., Content-Type: text/html; charset=iso-8859-1 last) with the Apache/2.4.52 (Ubuntu) Server at <host> Port <port> footer, where host and port come from the request's Host header. Request handling follows
    Ubuntu's Apache defaults: TRACE is 405 (TraceEnable Off); an unknown method is 501; PUT/DELETE/PROPFIND
    on a static or missing path is 405; OPTIONS on a static or missing path is 200 with Allow and no PHP headers;
    PHP pages answer every method except TRACE, as mod_php does; a malformed request line, bare-LF line endings,
    HTTP/1.1 without Host, or an invalid Host is 400 and the bad input is not echoed; a request line over 8190
    bytes is 414; oversized or too many header fields are 400 with Apache's note; HTTP/0.9 GET gets the body only.
    Keep-alive works as on Ubuntu (5 s, at most 100 requests per connection, Keep-Alive only when the client asks).
    Static bait files are served without PHP headers and with Last-Modified, ETag and Accept-Ranges.
    Malformed requests and unusual methods raise http.scanner.
  • Port scan detection on small decoys: a source touching every port of one decoy within 10 minutes (roles with at
    least 3 ports: the 4-port file server, the web role) raises scan.detected, besides the existing rule (5 distinct
    decoy ports in 10 minutes). A full scan of the file-server decoy was not detected before. SMB plus RDP from one
    client (3 of the file server's 4 ports) is not a scan.
  • Docs: README "Community Edition limits" and ATT&CK coverage; INSTALL §6 is now "alerts.json, events.json and Export";
    the lab guide uses tests/lab_scenarios.sh.

Removed

  • Telegram and JSON webhook alerting, the hourly alert cap, the alert delivery log (alerts table) and GET /api/alerts.
    rpot.json keys alerts.telegram, alerts.webhook and alerts.public_url.
  • Offline spool replay: events from a manager outage are no longer queued and replayed (RPOT_SPOOL_MAX_MB is gone).
  • Decoys on extra addresses: macvlan networks, new free IPs, VLAN placement, vendor MAC addresses and the free-IP
    check. Installer flags --placement, --decoy-ips, --company, --nics; rpotctl flags --company, --netbios,
    --email-domain, --user-format, --macvlan, --gateway, --host-network, --decoy-net, --own-ip and
    --net-mode macvlan; rpot.json keys decoy_nets and hostnet.
  • rpotctl persona and rpotctl setup: the persona and the baits cannot be changed after install.
  • --roles all (5 roles never fit the limit).
  • Upgrade code for v0.3 databases and configurations.

Security

  • SMB IPC$ is pipe-only (pentest F-01, critical): before, an anonymous session could list and download the
    decoy's working directory (the ReD-Pot source) through IPC$. Now only the srvsvc/wkssvc pipes open there;
    any other name is "not found", the share root is "access denied" and a listing is refused. IPC$ is backed by an
    empty directory, ADMIN$/C$ are never mapped to a directory, the decoy process runs from an empty directory,
    and a share table that maps outside the bait tree stops the decoy. Every attempt to use IPC$, ADMIN$ or C$
    as a disk is a new level-3 event smb.admin_share ("Tried to read admin share …").
  • Decoy interfaces show only the decoy: on each decoy interface only that decoy's ports answer; every other port
    (IPv6 included) looks closed, and the interface answers ARP only for its own address. This also hides other host
    services on a decoy interface, including the management interface when a decoy shares it. Admin ports (web UI,
    sensor API, host SSH) are accepted only on the management interface.
  • When a decoy shares the management interface, the web UI, sensor API and host SSH allow-lists default to the
    admin's own SSH client (/32), and --yes requires --allow-ui, --allow-sensor and --allow-ssh. An admin
    connected over IPv6 keeps SSH access through a /128 rule.
  • The same for remote sensors: when the decoy is on the interface that reaches the manager, the moved host SSH answers
    only the SSH allow-list (default: your own SSH client) and the current session; --yes requires --allow-ssh.
    Before, it was open to the whole decoy LAN and --allow-ssh was ignored on sensors.
  • The sensor join string (a one-time token) is no longer put on any command line during the install: the enrollment
    container receives it in its environment. Give it to the installer by pasting it when asked, on standard input
    (--join -) or in a root-only file (--join-file FILE, mode 600). --join 'redpot://…' still works, but warns that
    local users can read it in the process list while the installer runs. The dashboard's Add a sensor now shows the
    command sudo ./install.sh sensor and, in its own copy box, the join string to paste; the API's command field no
    longer contains the join string.
  • When a decoy shares the management interface, the installer first lists the host services that listen there and
    will become unreachable on that interface (for example a VPN server), and asks for confirmation. With --yes it stops
    unless --close-other-services is given.
  • The server does not route for a decoy LAN. Forwarded traffic arriving on a decoy interface is refused, and traffic
    into a Docker network is accepted only for published ports, so a direct route to the console container's address
    (port 8740/8741) cannot bypass the allow-lists.
  • The firewall fails closed: firewall.sh apply checks that every chain is hooked (iptables, ip6tables while IPv6 is on,
    DOCKER-USER or FORWARD). If a chain is missing, it exits with an error that names the chain and the rules are
    removed again. With Docker's nftables firewall backend the chain is hooked into FORWARD directly. The self-test
    reports whether the firewall is in effect.
    • Install: the installer first checks these prerequisites, before it stops services or moves SSH. If apply still
      fails, the installer stops and the decoys are not started. The check tests the egress guard's owner match in a
      temporary chain hooked from OUTPUT and the FORWARD hook with an empty chain (the kernel accepts the owner match
      only on the output path; mixing both made every install with a decoy fail), and removes its chains again.
    • add-decoy / remove-decoy: if the new rules fail, the previous rules are put back for the running console and
      decoys and the change is not started; if that fails too, the whole stack is stopped.
    • The redpot service first stops its containers, then applies the firewall, then starts them. After an unclean
      shutdown Docker restarts the containers on its own when it starts, so they run without the firewall until the
      service starts (after the network is online); if apply fails then, they are stopped and stay stopped.
  • The NetBIOS name-service exception of the egress guard allows only small UDP answers from port 137, out of the decoy's
    own interface to a host in that interface's subnet, never to port 53.
  • The installer refuses to install when uid 10001 (the containers' user) is used by a host account or host process.
  • add-decoy next to a running domain controller decoy works: a socket bound to one interface (the DC's NetBIOS socket,
    0.0.0.0%eth0 in ss) counts only for that interface's addresses when the installer checks for busy ports.
  • Egress guard: decoy processes in the host network may only reply and report to the sensor API; the web UI,
    127.0.0.1, the LAN and the internet are refused.
  • SMB share enumeration (net view, smbclient -L) works on host-mode decoys behind the egress guard: the
    srvsvc/wkssvc named pipes reach their RPC handlers inside the decoy process instead of over TCP 127.0.0.1,
    so the guard keeps loopback closed and no RPC port listens on 127.0.0.1 any more.
  • SMB srvsvc server info (NetrServerGetInfo) answers with the decoy's name, persona OS version and server type
    (domain controller bits on a DC); it used to fail and drop the pipe.
  • The self-test lists host services that listen on all addresses.
  • The console makes no outbound connections for alerting, so network monitoring sees no alert traffic from it.
  • The undelivered log is opened write-only with O_NOFOLLOW, never read back, and capped.
  • Enrollment validates the decoy IP; slot checks run inside the fleet lock, so concurrent join requests cannot exceed
    the limit. A sensor beyond the limit gets 403 on ingest and config.
  • The export requires an admin session and the CSRF header, is sent with Cache-Control: no-store, and contains bait
    user names and attempted passwords: keep exported files private.
  • The image has no setuid/setgid files. Local host-network decoys and the decoy of a remote sensor now also run with
    no-new-privileges:
    they start as root with only NET_BIND_SERVICE, SETUID, SETGID and SETPCAP, and the entrypoint drops at once
    to uid 10001 with setpriv, cutting the bounding set to NET_BIND_SERVICE and keeping only it (ambient) for ports
    22/445. The console, sensor and relay roles refuse to run as root without that drop.
  • sensor.gap notes are rate-limited: at most one per decoy per 5 minutes (later reports in that window are merged
    into it, counts summed), and decoy-incident notes keep only their newest 20 lines / 4 KB. The first/last timestamps
    are the decoy's own clock and are dropped when older than 30 days or in the future.
  • A remote sensor that follows its interface's address never listens on 0.0.0.0 when that interface has no IPv4
    address: it waits briefly, then exits and is restarted.

Container image

ghcr.io/redsidesecurity/red-pot:1.0.0 (linux/amd64, linux/arm64)

  • digest: sha256:277f64452e99515fb6fdd9c19587b20acd5d44eef2928c9bb6b3f29d825dbe6d
  • source hash label io.redside.src-sha256: c1ade7253a657b8191385e174640c3665cf91ffe3a50ae11d8b768977c43c5a8
  • provenance: gh attestation verify oci://ghcr.io/redsidesecurity/red-pot:1.0.0 --repo RedSideSecurity/ReD-Pot