Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore

Odin Research / What is / C2 and post-exploitation

What Is MITRE Caldera? How the Adversary Emulation Platform Works

An open-source platform that runs ATT&CK-style attacks for red and purple teams: how it works, why it floods C2 feeds, and how to tell a lab from a threat.

By
Odin by Brandefense
Updated
21 Nov 2026
Reading time
8 min read
Type
Adversary emulation platform

In short

MITRE Caldera is an open-source adversary emulation platform built on the MITRE ATT&CK framework. A central server sends tasks to agents such as Sandcat, which run attacker techniques on test machines. Red and purple teams use it to check defenses. Its agents behave like real implants, so defenders must tell sanctioned labs from hostile use.

Key takeaways

What it is
An open-source adversary emulation platform, first built by MITRE and now contributed to the Apache Incubator as Apache Caldera. It automates ATT&CK-style attacks for testing.
How it works
A Python server with a web interface and REST API tasks agents. The default agent, Sandcat, is written in Go and beacons to the server over HTTP.
Why it matters
Caldera agents act like real implants, and an exposed server is a command channel into every enrolled host. A 2025 flaw, CVE-2025-27364, allowed remote code execution on the server.
What Odin Research saw
257 Caldera endpoints in Monograph 22, but 239 were banner look-alikes in one security vendor's address space. No Caldera address was published for blocking.
What to do first
Register every sanctioned Caldera server, keep it off the internet, patch it, and alert on Sandcat traffic from hosts that are not part of an approved test.

What is MITRE Caldera?

MITRE Caldera is an open-source adversary emulation platform built on the MITRE ATT&CK framework. It runs real attacker techniques through agents on test machines, so security teams can see which steps their defenses catch. Red teams, purple teams and researchers use it, and its agents behave much like hostile implants.

MITRE developed Caldera with support from the US National Science Foundation. MITRE's project site describes nearly a decade of development. On 20 May 2026, MITRE announced that it was contributing the project to the Apache Incubator as Apache Caldera (Incubating). Repositories, documentation and release processes are moving to Apache Software Foundation infrastructure.

MITRE lists four main uses: red teaming, purple teaming, security validation and cyber research. In each case the goal is the same. The team picks an adversary profile, Caldera runs its techniques against enrolled hosts, and analysts check what their EDR, SIEM and SOC saw.

Because its techniques map to MITRE ATT&CK, results are easy to compare with threat reports. That same realism is why a Caldera server outside a known lab deserves a closer look.

How MITRE Caldera works

According to the project's README, Caldera has two parts. The core is an asynchronous command-and-control (C2) server with a REST API and a web interface. Plugins then add agents, reports, technique collections and other features.

The main building blocks are:

  • Agents. Sandcat is the default agent. The plugin documentation says it is written in Go for cross-platform use and beacons to the server over HTTP to receive instructions. Manx adds a TCP-based reverse shell.
  • Abilities and adversaries. The Stockpile plugin supplies abilities (single techniques), adversary profiles, a planner and facts. The Atomic plugin imports Red Canary's Atomic tests.
  • Planners. A planner decides the next step from facts the agents collect, such as user names or host names. This is what makes Caldera an automated emulation tool rather than a simple remote shell.
  • Optional channels. Sandcat extensions can use GitHub gists as a C2 channel, or relay traffic peer to peer over HTTP or SMB named pipes.

Monograph 22 describes the Sandcat check-in on the wire. The target sends POST /file/download with a file: sandcat.go header. The server compiles a fresh Go binary that embeds the server address and group. The agent then sends Base64-encoded JSON, which starts with eyJ, to POST /beacon and receives its next task.

MITRE Caldera requirements, versions and security history

The README lists a Linux or macOS system with Python 3.11 or later and pip3. It recommends 8 GB of RAM, two CPUs, Go 1.24 or later for compiling Go agents, and Node.js 16 or later for the version 5 web interface. The server listens on port 8888 by default. The SSL plugin adds HTTPS on port 8443 through HAProxy.

The README gives a default login and strongly recommends running Caldera in a secure environment, not exposed to the internet. That advice matters. A Caldera server can task every agent that has checked in.

In February 2025, NVD published CVE-2025-27364. It is a remote code execution flaw in Caldera's dynamic agent compilation. A crafted request to the API that compiles and downloads Sandcat or Manx agents could run arbitrary code on the server. It affects Caldera through 4.2.0 and 5.0.0 before a specific fix commit. The CVSS 3.1 score from the CNA is 10.0. The README points users to version 5.1.0 or later for the patch.

The repository had about 7,500 stars and 1,400 forks on GitHub on 21 November 2026.

Why MITRE Caldera matters to defenders

Caldera is a legitimate testing tool, and nothing in the sources reviewed for this page documents a named threat group using it in an intrusion. It still matters to defenders for three reasons.

  1. Its agents look like implants. Sandcat beacons over HTTP, runs commands through PowerShell, cmd or sh, and can chain peer-to-peer. An unexplained Sandcat process deserves the same response as unknown remote access malware until someone confirms an approved test.
  2. An exposed server is a control point. Anyone who reaches an internet-facing Caldera console with weak credentials, or an unpatched one with CVE-2025-27364, gets a ready-made C2 into enrolled hosts. See How to respond to misconfigurations.
  3. It pollutes threat feeds. Caldera banners and TLS fingerprints are easy to match, so feeds list many lab and look-alike servers as C2. Blocking them wholesale breaks business traffic and hides real threats.

What Odin Research observed

Odin Research has looked at Caldera in two studies. Both reached the same conclusion: most servers labeled Caldera are not attack infrastructure.

  • A top-ten label inflated by a vendor lab. In Monograph 02, Caldera ranked ninth by raw volume with 240 records, on ports 443 (118), 80 (114), 16018 (5) and 8080 (3). Routing data put 239 of them (99.6%) inside one security vendor's network, consistent with a Fortinet demo lab.
  • A shared TLS fingerprint. The same report found one JARM value on 90 vendor-lab Caldera records across 85 hosts, and also on Brute Ratel C4 records. JARM measures the TLS stack behind a proxy, not the program, so it should never trigger an automatic block.
  • 257 endpoints, none confirmed hostile. Monograph 22 counted 257 Caldera endpoints on 135 hosts. Of these, 239 were banner look-alikes on 117 US hosts in Fortinet corporate address space, a fingerprint collision rather than Caldera activity.
  • 18 cloud servers on watch. The other 18 endpoints sat on 13 AWS, 2 DigitalOcean, 2 Alibaba and 1 TT DOTCOM hosts. Fourteen used the default port 8888. Any of them could be a sanctioned purple-team lab, so no Caldera address was published.

The lesson: treat Caldera listeners as hunting leads and watchlist entries. Act only when a second, independent signal shows hostile use.

How to detect MITRE Caldera

Look for the agent, not just the server banner:

  • Agent downloads. An outbound POST /file/download request with sandcat.go in the headers is a Sandcat download.
  • Beacons. Repeated POST /beacon requests whose body starts with eyJ match Sandcat's Base64 JSON check-ins. HTTPS traffic needs TLS inspection to see this.
  • Command lines. Process creation with -server http and -group, or that names sandcat.go, 54ndc47 or manx.go, points to a Caldera agent.
  • Port alone is weak. Port 8888 and generic web banners are shared by many tools. Join them with process data before you escalate.

Monograph 22 publishes ready-to-use rules for these signals. They include Suricata signatures for Sandcat downloads and check-ins, a Sigma rule for Sandcat and Manx execution, a YARA rule for Caldera agent files and a Microsoft Sentinel KQL hunt. Each rule carries an allowlist for sanctioned lab hosts. All rules are in the Defense Kit. For background on rule formats, see YARA and Sigma rules.

How to run MITRE Caldera safely and respond to unknown agents

  1. Keep a register. List every sanctioned Caldera, Empire, Viper and Metasploit host. Add them to SIEM allowlists by exact hostname, so a real alert is never hidden by a broad exclusion.
  2. Hide the console. Put Caldera behind a VPN or mutual TLS, change the default credentials and never expose port 8888 to the internet.
  3. Patch. Run version 5.1.0 or later to close CVE-2025-27364, and follow the project's move to Apache release channels.
  4. Scope every operation. Record the start and end time, the agent group and the target hosts of each test. Any Sandcat activity outside that window is an incident until proven otherwise.
  5. Do not auto-block feed lists. Filter vendor look-alikes and lab ranges first, then hunt to confirm.

MITRE Caldera compared with other team C2 frameworks

The figures below come from Monograph 22, which mapped ten team C2 and emulation frameworks in the same two-week window.

FrameworkWhat it isEndpoints (hosts)Confirmed hostile
MITRE CalderaAutomated emulation planner with Sandcat and Manx agents257 (135)0
StarkillerMulti-operator console for PowerShell Empire5 (5)0
MetasploitExploit and C2 framework75 (65)17
SuperShellBrowser reverse-SSH console180 (180)2
AdaptixC2Plugin-based teamserver101 (89)0

Caldera had the largest raw count and no confirmed hostile servers. That gap is typical of dual-use emulation tools. For hosted alternatives in the same family, see our pages on Viper and Sliver.

Frequently asked questions about MITRE Caldera

What is MITRE Caldera?

MITRE Caldera is an open-source adversary emulation platform built on MITRE ATT&CK. A central server tasks agents such as Sandcat to run attacker techniques on test hosts, so red and purple teams can check what their defenses detect.

Is MITRE Caldera malware?

No. Caldera is a legitimate, Apache-2.0-licensed security testing tool. Its agents behave like implants, though, so an unexplained Sandcat or Manx agent should be handled as a possible intrusion until an approved test is confirmed.

Is Caldera free and legal to use?

Yes. Caldera is free and open source. Using it is legal on systems you own or are authorized to test. Running its agents on systems without permission is illegal in most countries.

Who made Caldera?

MITRE developed Caldera with support from the US National Science Foundation. In May 2026, MITRE announced it was contributing the project to the Apache Incubator, where it continues as Apache Caldera (Incubating).

What is the Caldera Sandcat agent?

Sandcat is Caldera's default agent. It is written in Go, runs on several operating systems and beacons to the Caldera server over HTTP to receive tasks. Manx is a second agent that provides a TCP reverse shell.

What port does Caldera use?

The Caldera web interface and API listen on port 8888 by default. The SSL plugin adds HTTPS on port 8443. Odin Research found that 14 of 18 cloud Caldera servers used port 8888.

How do I detect Caldera in my network?

Look for POST /file/download requests naming sandcat.go, POST /beacon requests with Base64 JSON bodies, and command lines with -server and -group or names such as sandcat.go and 54ndc47. Odin Research publishes Suricata, Sigma, YARA and KQL rules for these signals.

Is there a vulnerability in Caldera?

Yes. CVE-2025-27364 is a remote code execution flaw in Caldera's dynamic agent compilation, rated CVSS 3.1 10.0 by the CNA. It affects Caldera through 4.2.0 and early 5.0.0 builds. The project points users to 5.1.0 or later.

Caldera vs Atomic Red Team: what is the difference?

Atomic Red Team is Red Canary's open-source collection of tests. Caldera is a full platform with a server, agents and planners that chain techniques together. Caldera can import Atomic tests through its Atomic plugin.

Sources

  1. MITRE, Caldera project site. caldera.mitre.org/ (checked 21 November 2026)
  2. Caldera repository and README, GitHub. github.com/mitre/caldera (checked 21 November 2026)
  3. Caldera documentation, Plugin library. caldera.readthedocs.io/en/latest/Plugin-library.html (checked 21 November 2026)
  4. NVD, CVE-2025-27364. nvd.nist.gov/vuln/detail/CVE-2025-27364 (checked 21 November 2026)

Related tools and terms

All Odin Research explanations · Brandefense glossary

Odin Research

Get the rules, not just the definition

Every Odin Research report ships with verified indicators and detection rules you can load today.

Open the Defense Kit