What our sandbox established
Use ARES as a lab evaluation candidate if campaign orchestration is your goal. Our evidence supports a successful installation and build-step outcome, followed by an unfinished test run. It does not support a client-deployment recommendation.
Our run on 2026-09-30 used commit 49bd11a, whose full SHA is 49bd11ab27dca63b4a7d911069eb9883fd70c36e. The stored sandbox configuration specifies 3 CPUs, 8 GB of memory and the ghcr.io/astral-sh/uv:python3.12-bookworm image. These observations belong to that sandbox run; no new execution was performed for this correction.
Our sandbox recorded installation success in 73 seconds, with 196 packages occupying 499 MB. The recorded build step returned zero in 9 seconds. The test command timed out after 900 seconds with exit code 124.
Our harness stores each command's success flag, elapsed time, exit code, timeout flag and available output tail. The build tail is empty, so its successful exit does not identify a tested application workflow. The testing-methodology page explains how to interpret these records. Its currently published procedure does not establish which methodology version governed this older run.
What the unfinished tests mean
The retained test output reached a 22% progress marker before the timeout. That marker is not a passing-test percentage or a measure of feature coverage. The tail includes skipped tests, shown as s; skips are not failures. No failure is visible in the supplied tail, but a partial tail cannot establish that the suite passed.
The attached record contains no aggregate parsed test count. We therefore cannot report a total of unique tests, reconcile a complete suite or infer how much longer it would take. The release page's claim of 1078 passing unit tests belongs to the upstream v6.0.0 release, not our September run.
The stored pip-audit result records one known advisory affecting one package. It does not identify the package, advisory or severity. Scan coverage is unknown because the record has no scannedPackages value. This advisory count is a triage item, not a security assessment of ARES.
Campaign controls need a verified execution boundary
The ARES repository describes campaign scope controls, attack-path planning, encrypted evidence storage and report generation. These capabilities explain why an experienced team might evaluate it. They remain documented capabilities here, rather than workflows demonstrated by our sandbox.
The current security model says Python socket hooks do not intercept external binaries or compiled extensions issuing direct kernel syscalls. That distinction matters when a module launches another executable: checking Python connections alone cannot establish containment of that executable.
The security model marks actual runtime verification of Linux Netfilter rules, Windows Defender Firewall rules, subprocess network namespaces and privilege dropping as incomplete or not verified under the stated conditions. It also explains that mocked command tests cannot prove physical packet filtering.
Strict mode rejects several configurations without the requested boundary unless the caller explicitly permits an exception. That is a documented policy, not evidence that our run exercised the corresponding kernel controls. The current document also warns that online Docker bridge networking does not restrict external binaries to campaign scope. Validate the exact operating system, isolation tier and network policy you intend to use.
Setup extends beyond installing Python packages
The supplied README lists Python 3.11 or 3.12 and Node.js 18 or newer, with separate frontend installation steps. Our Python installation result should not be read as completion of that entire development path.
For a deployment, generate separate signing and encryption secrets and back up the encryption key securely. The current security model says changing the bootstrap password environment variable after the initial account exists does not reset that account's password. Use the supported password-change workflow for an existing account.
Before an authorized exercise, finish the suite in your environment, identify and assess the recorded advisory, and verify both blocking and cleanup of the chosen network boundary. Keep engagement evidence and credentials within the deployment's access and backup policy. These are next steps for evaluation, not tasks our sandbox completed.
Choose an alternative by the work you need
Apache CALDERA is an alternative to evaluate for adversary emulation with a core server and plugins. Its README describes a command-and-control server, REST API, web interface and plugins for agents, reporting and technique collections. This establishes a comparable orchestration role; it does not establish that CALDERA is safer or easier to operate. Its own deployment guidance recommends a secure environment rather than internet exposure.
Atomic Red Team is a library of tests mapped to MITRE ATT&CK that can be executed directly from the command line. Choose it when your immediate task is to review and run individual detection tests. It can also complement a broader campaign workflow; adopting a test library is a different decision from adopting an orchestration platform.
How to read the ratings
These editorial ratings use a 1 to 5 scale: 1 means substantial obstacles, 2 means material work or unresolved validation, 3 means adequate evidence for the dimension, 4 means strong evidence with stated limits, and 5 means unusually complete evidence within the assessed scope. They are not security certifications.
Setup ease is 2 because installation succeeded but the documented frontend and host-boundary work remain substantial. Documentation is 5 for the supplied security model's explicit policies, exceptions and verification limits; this rates the explanation, not the implementation's safety. Maturity is 2 because our suite remained unfinished and deployment behavior was not established. Community is unassessed because the supplied independent extracts do not establish participation or support responsiveness.

