A source-built, locally bound distribution of ARTEX, an open-source multi-agent security testing harness — with an English interface, a hardened supply chain, and deployment controls.
This repository exists because the prebuilt tomicdd/artex Docker image was
found to contain an unauthenticated command-execution backdoor in November 2026.
This build is compiled from the audited v0.3.15 source tag and keeps upstream's
full runtime capabilities, including Node 22, Playwright, Chromium, and nmap.
The compose file binds the console to localhost and does not publish the
recording proxy.
Authorized testing only. ARTEX runs an autonomous agent that can execute shell commands and send network traffic. Use it only against systems you own or have explicit written permission to test.
This distribution is maintained by FailSafe Cyber Frontier Labs, the team behind GlassBreak and FailSafe's managed security-testing runtime. We review upstream changes, maintain the hardened build, and operate ARTEX for teams that need the capability without building the guardrails themselves.
- General and security questions: hello@getfailsafe.com
- Managed ARTEX runs and cyber-model access: hello@getfailsafe.com
- Website: https://getfailsafe.com
ARTEX is an agentic harness: the model plans, calls tools, and reasons about targets across many turns. For authorized security work we recommend a cyber-specialized model — one trained or fine-tuned for security tasks (cyber-domain SFT) with minimal refusal behavior, rather than a general-purpose chat model. General consumer models routinely refuse reconnaissance, payload construction, or exploitation reasoning even when the engagement is authorized, which turns a legitimate test into a stream of blocked tool calls and wasted tokens.
GlassBreak is FailSafe's cyber-grade model, built for this workload. It is what we run when we operate ARTEX ourselves, and it is what the managed runs use. If you do not already have a cyber model you can use, contact hello@getfailsafe.com to arrange a managed run or rent model access for runs on infrastructure you control.
- Autonomous multi-agent testing. A goals agent decomposes the objective, an event-driven planner explores, and concurrent workers execute intents. New assets, facts, findings, and results feed back into planning. A main conversation agent lets a human steer mid-run.
- Real execution capabilities. Shell and file tools, browser automation
through Playwright MCP/CLI and bundled Chromium, skills, MCP servers, custom
Python tools, DNS/network recon, and raw SYN scans with
nmap. - Evidence-backed findings. Findings keep task/asset lineage and recorded
HTTP evidence, generate reports, and enter a dedicated retest workflow with
reproduced,fixed, orinconclusiveverdicts. - Traffic capture. A recording proxy indexes HTTP exchanges, bodies, and archives; evidence binds to findings; archives export and reclaim safely.
- Operational controls. Admin authentication, configurable tool intercept/allow/deny rules, optional model-judge approvals, pause/resume, queueing, cancellation, deadlines, restart recovery, and task isolation.
- Full English interface and documentation. The console, agent skills, and project documentation are in English for international teams.
- Fast static delivery. The embedded web console is served with Brotli/gzip
compression, content-hash ETags, and immutable caching, so the UI stays
responsive when the binary is reached directly on
:8787— no reverse proxy required for acceptable load times.
| Area | Upstream default | This build |
|---|---|---|
| Image | prebuilt tomicdd/artex (backdoored) |
compiled from this source tree |
| Runtime user | root | root (upstream parity) |
| Console binding | 0.0.0.0:8787 |
127.0.0.1:8787 |
| Recording proxy | published on 8788 |
not published |
| Static UI delivery | uncompressed, no cache validators | Brotli/gzip + immutable cache + ETags |
| Interface language | Chinese | English |
| Browser automation | Node 22 + Playwright + Chromium | same, pinned versions |
| Browser MCP launcher | npx @playwright/mcp |
pinned playwright-mcp |
| API-recon dependencies | installed during a run | preinstalled from a locked official registry |
| Online self-update | enabled, points upstream | compiled off; no remote check or download |
| Capabilities | default Docker set | default Docker set + no-new-privileges |
| Base images | floating tags | pinned by digest |
The API, UI, skills, browser MCP, and browser-driven recon are otherwise
unchanged. Node, Playwright, the Playwright MCP/CLI, and Chromium are
preinstalled; nmap raw scans and runtime apt/npm -g installs work because
the container runs as root, as upstream does. The deployment hardening that
remains is the loopback-only console, an unpublished proxy, no-new-privileges,
a persisted JWT key, and digest-pinned base images.
FailSafe can host and run ARTEX safely for you using GlassBreak, our cyber-grade model. We handle the isolated runtime, scoping, evidence handling, and the operational controls that are easy to get wrong in a self-hosted deployment. We can also rent cyber-model access (including GlassBreak) for runs on infrastructure you control.
If you want a managed run or model access, contact us:
- Email: hello@getfailsafe.com
- Contact form: https://getfailsafe.com/contact-us
Requirements: Docker Engine with Compose v2, and about 5 GB of free disk
(the browser stack and Chromium account for most of the image size).
If your Docker installation only provides the standalone docker-compose
command, substitute it for docker compose below; make detects either.
cp .env.safe.example .env.safe
chmod 600 .env.safe
# edit .env.safe: set POSTGRES_PASSWORD and, optionally, an LLM key
docker compose --env-file .env.safe -f docker-compose.safe.yml up -d --buildOpen http://127.0.0.1:8787 and complete the first-run setup. If you left the LLM key blank, configure a provider in the console before starting a scan.
Check status and logs:
docker compose --env-file .env.safe -f docker-compose.safe.yml ps
docker compose --env-file .env.safe -f docker-compose.safe.yml logs -f artexStop without deleting data:
docker compose --env-file .env.safe -f docker-compose.safe.yml downDelete everything, including the database and ARTEX data:
docker compose --env-file .env.safe -f docker-compose.safe.yml down -vThe compose file publishes the console on 127.0.0.1 only. To reach it from
your laptop, use an SSH tunnel:
ssh -L 8787:127.0.0.1:8787 user@hostIf you put a reverse proxy in front, forward both / and /api/ to the
backend; the UI's event streams use same-origin URLs in production, so no
extra port needs to be exposed.
cd web
npm ci --ignore-scripts
npm run build:static
cd ..
rm -rf server/webui/dist
mkdir -p server/webui/dist
cp -R web/out/. server/webui/dist/
CGO_ENABLED=0 go build \
-trimpath -buildvcs=false -tags embedui \
-ldflags="-buildid= -X main.version=v0.3.15-failsafe.3" \
-o artex ./cmd/artexThen run ./artex -addr 127.0.0.1:8787 -proxy 127.0.0.1:8788 with
ARTEX_PG_DSN set to a PostgreSQL database you control. This bare-binary path
does not install Node, Playwright, Chromium, or the recon CLI tools; use the
Docker image for full upstream capability.
make verifyThis checks the known implant indicators (verifyCaptcha, tomicdd,
congdonglinux), prints the binary SHA-256, and lists the image ID. See
audit/artex-v0.3.15-audit.md for the full
audit of this tag.
Every push and pull request runs:
go vet ./...and every Go test in its own process against a fresh PostgreSQL database (ci/go-tests-isolated.py);- the web console's production build (
npm ci --ignore-scripts+npm run build:static); - an end-to-end check of the single binary with the embedded UI
(
ci/e2e_binary.py): static export, SPA fallback, missing-asset 404, Brotli/gzip, ETag304, immutable cache headers, authentication, task creation, disabled updater, and graceful shutdown; - focused PostgreSQL regressions, the full image build,
make verify, the preinstalled api-recon runtime, browser/MCP discovery, and a boot smoke test on an internal network.
Run the local checks yourself:
go vet ./...
ARTEX_PG_DSN=postgres://user:pass@127.0.0.1:5432/artex_test?sslmode=disable go test ./...Unit tests for the static UI server (compression negotiation, ETag
revalidation, cache policy, Range/HEAD handling) run as part of go test ./....
.env.safecontains the database password and optional LLM keys. Keep it0600and out of version control.- The database lives in the
pgdatavolume; ARTEX data and evidence live inartex_data; the JWT signing key lives inartex_key. Back up all three. - ARTEX stores scan evidence, which may include credentials, tokens, and
customer data. Treat
/app/dataand the database as sensitive. - For a backup, stop the stack and archive the volumes, or use
pg_dumpinside thepostgrescontainer.
Read UPSTREAM.md first. In short: review the upstream diff,
rebuild, run make verify, update the audit report, and only then deploy.
ARTEX is AGPL-3.0. See LICENSE and UPSTREAM.md.
This fork's modifications are released under the same license. The upstream
README is preserved (in English) as README.upstream.md.
