What ARTEX Looks Like From a Decoy
We joined a controlled ARTEX WebFetch with threat.gg model-plane telemetry to show what the framework reveals on both sides of an agent run.
AI agents are changing the speed and scale of attacks. The HTTP requests they send still look surprisingly familiar.
We saw that firsthand when we ran ARTEX against webdecoy.com in a controlled test. ARTEX used its WebFetch tool to request the root page. Our edge sensor recorded this User-Agent:
norma/0.4WebDecoy detected the request as Norma WebFetch, categorized it as a scraper, and assigned a threat score of 16. The new framework-attribution layer independently labeled the same detection:
Framework: ARTEX
Attribution: Confirmed test
Rule: artex.signed-test.v1
Observed client: Norma WebFetchThat separation is the important part. The request was easy to detect. Proving which framework launched it required different evidence.
Why ARTEX matters
ARTEX is an open-source, agentic penetration-testing framework. Its agents can discover targets, choose tools, test vulnerabilities, and pursue an objective with limited operator input.
Check Point Research found 2,557 systems that had hosted ARTEX panels, including 1,608 live, publicly accessible instances at the time of its research. Data exposed by ten misconfigured deployments showed operators setting objectives and allowing agents to perform much of the work autonomously. The observed campaigns included weak-password attacks, authorization bypass, remote code execution, credential theft, and supply-chain compromise. Read the Check Point research.
The underlying exploitation techniques remain recognizable. The change is orchestration. A framework can run more paths in parallel, adapt after each result, and continue with far less attention from an operator.
That creates a practical question for defenders: what does ARTEX look like from the target’s side?
What the target can observe
ARTEX uses several tools, so there is no universal ARTEX network fingerprint. A command executed through Bash can inherit the signature of curl, nmap, nuclei, or another utility. WebFetch traffic is different. ARTEX’s WebFetch tool uses the Norma SDK, whose default User-Agent is norma/0.4. The ARTEX defensive documentation describes this target-visible behavior directly. See the ARTEX defense notes.
ARTEX also has a framework-specific enrichment client that sends a User-Agent beginning with:
artex-enrich/These strings are useful detection signals, but they do not carry the same evidentiary weight.
artex-enrich/ is specific to ARTEX, although an operator can change or spoof it. norma/0.4 identifies Norma WebFetch, a client library that can be run independently or by other software. A Norma request alone cannot establish that ARTEX sent it.
Static User-Agent signatures also have an obvious limit: an operator can edit them. Their absence does not establish safety, and their presence does not establish operator identity.
What the model endpoint can observe
The target sees the HTTP clients ARTEX chooses. A model endpoint sees a different layer: the instructions and tool schemas that drive those choices.
threat.gg shared 4,342 request records collected by its Ollama/AI honeypot network on November 9, 2026. WebDecoy analyzed the corpus independently. We did not ingest raw prompts or target material into the product; the detection rule retains only registered indicator IDs and SHA-256 fingerprints.
The corpus had a stable shape:
- 4,292 requests contained the same ARTEX prompt-contract family, including its continuous-engagement authorization and its instruction to treat target stop messages as data.
- 3,917 requests exposed tool-enabled agent work.
- 3,017 carried the canonical worker schema, including
search_all_worker_traces,get_worker_trace,node_detail, andfleet_run. - 892 carried the planner/controller schema, with tools such as
graph_overview,add_intent,prove_goal, andsteer_work.
The trace also shows why target-side detection cannot depend on one framework client. Across deduplicated conversation histories, ARTEX invoked Bash 1,491 times, listed directories 1,124 times, read files 224 times, and called WebFetch once. We found 227 curl-bearing Bash commands. Most of those curl commands did not set an explicit User-Agent.
Those tool-call counts describe deduplicated conversation histories, not 1,491 independently observed network requests. They show the framework’s choices at the model layer; they do not prove what reached any particular target.
The strongest model-plane fingerprint was not omp/18.2.11, the GLM model name, or curl. Oh My Pi is a general-purpose agent runtime, models are shared infrastructure, and curl is ubiquitous. The useful evidence was the conjunction of a stable ARTEX prompt contract and a role-coherent ARTEX orchestration schema in the same request. ARTEX’s public changelog independently documents distinctive tools such as search_all_worker_traces and get_worker_trace. See the ARTEX changelog.
This gives defenders two complementary observation planes:
| Observation plane | Evidence | Defensible conclusion |
|---|---|---|
| Model API, observed by threat.gg | ARTEX prompt markers plus a role-coherent orchestration tool schema | Likely ARTEX framework activity |
| Target HTTP, observed in WebDecoy’s controlled test | norma/0.4 fetching the approved page | Norma WebFetch reached the target |
| Signed WebDecoy test relay | Valid property-scoped, expiring provenance | Confirmed ARTEX controlled test |
The threat.gg traffic and our controlled WebFetch are separate observations. They cannot be joined by actor, address, operator, or run. They corroborate what ARTEX exposes on each side of its workflow without claiming that the same party generated both datasets.
The test
We ran ARTEX on an isolated Mac mini and gave it one objective:
Use WebFetch exactly once to fetch the approved relay URL and then stop.
Do not use any other tool or URL.The relay accepted only GET /, forwarded to https://webdecoy.com/, and enforced a small hard cap on upstream requests. The target was a static site with no application surface for ARTEX to exploit. The exercise measured detection and attribution only.
For provenance, an authenticated WebDecoy user created a short-lived test token scoped to the organization, property, framework, and run. The edge sensor relayed that token with the request. Ingest verified its HMAC signature, expiry, and property scope, stored the decoded provenance, and discarded the bearer token.
The successful request reached webdecoy.com, returned HTTP 200, and produced an edge-worker detection at 16:57 UTC on November 14, 2026.
WebDecoy kept four conclusions separate:
| Question | Result |
|---|---|
| Which client reached the site? | Norma WebFetch |
| Which framework launched it? | ARTEX, confirmed controlled test |
| How suspicious was this request on its own? | Threat score 16, minimal |
| What response was taken? | Logged and allowed |
The framework attribution did not increase the threat score or create a blocking rule. A controlled test proves provenance. It does not make the request malicious.
Why one header was not enough
Our first implementation could recognize artex-enrich/ and norma/. That solved client detection, but it left an attribution problem.
Labeling every norma/0.4 request as ARTEX would create false claims. It would also make response policy unsafe. Blocking a shared client fingerprint can lock out unrelated software that uses the same library.
WebDecoy now represents framework attribution as a separate correlation record with three tiers:
| Tier | Evidence required | What it means |
|---|---|---|
| Associated | One framework-specific but spoofable clue, such as artex-enrich/ | ARTEX-associated traffic was observed |
| Likely | A verified model-API observation combines ARTEX prompt markers with at least four role-coherent tools; or ARTEX enrichment and Norma WebFetch appear within ten minutes under a strong identity | Independent evidence families support the ARTEX attribution |
| Confirmed test | A valid, expiring, property-scoped token accompanies an authenticated controlled run | This request belongs to the named WebDecoy test |
Network proximity alone is insufficient for the Likely tier. Shared IP addresses, coarse ASN groupings, and cloud egress can combine unrelated clients. The correlation worker requires a stronger actor or a same-IP match reinforced by cryptographic, device, or JA4 evidence.
Every attribution records the rule that produced it, an explanation, and links to supporting detections where correlation is involved. Analysts can inspect the path instead of accepting a label with no proof.
The model-plane rule is deliberately one-way. It can raise attribution to Likely, but it does not change the threat score or create an enforcement rule. The reporter must use the authenticated SDK route, unknown indicators are rejected, and raw prompts, completions, tool arguments, request bodies, and target names are outside the accepted schema.
Detection, attribution, scoring, and response answer different questions
Security products often compress several decisions into one badge. That makes a dashboard look decisive, but it hides uncertainty.
WebDecoy treats these as separate layers:
- Detection records the request and the client directly observed by the sensor.
- Attribution connects that client to a larger framework when independent evidence supports the connection.
- Threat scoring estimates the risk of the observed behavior.
- Response decides whether to log, challenge, rate-limit, or block.
The distinction matters in this test. Norma was directly observed. ARTEX was established through signed provenance. The request was low risk because it fetched one public page. The sensor allowed it.
In an uncontrolled production event, the evidence may support only Associated or Likely. The interface says so.
What defenders can use today
For network and edge telemetry, start with two narrow signatures:
User-Agent starts with artex-enrich/
User-Agent starts with norma/Treat the first as ARTEX-associated. Treat the second as Norma WebFetch. Preserve the raw request context, timestamp, property, actor evidence, TLS fingerprint, and any decoy interactions so a correlation layer can evaluate the sequence.
Avoid making an enforcement decision from either User-Agent alone. Higher-confidence response should come from behavior such as credential-file requests, exploit paths, decoy hits, repeated enumeration, or a strong actor that persists across requests.
ARTEX’s source and dependencies continue to change. Detection based on a single static string will age quickly. Evidence correlation gives defenders a model that can absorb new clients and framework-specific signals without rewriting history or pretending every clue is proof.
What we learned
The controlled request was detectable from one header. The truthful ARTEX label required provenance. The threat.gg corpus added a second lesson: the orchestration layer has a fingerprint too, but it lives at the model endpoint rather than the target.
That is the broader lesson for agent detection. A target usually sees the tools an agent chose. A model endpoint can see the prompt contract and tool vocabulary that made the choice. Framework identity gets stronger when those independent planes agree, provided the product preserves their limits and does not invent a shared operator.
WebDecoy’s framework attribution currently supports ARTEX and reports Associated, Likely, or Confirmed test directly in detection details. The observed client remains visible beside it.
Read the framework-attribution documentation.
Ready to see what automated clients are reaching your site? Start free with WebDecoy.
Sources
Frequently Asked Questions
Does a norma/0.4 User-Agent prove that a request came from ARTEX? +
No. It identifies Norma WebFetch, a client library that ARTEX uses but other software can run independently. ARTEX attribution requires additional framework-specific, correlated or signed test evidence.
How does WebDecoy attribute traffic to ARTEX? +
WebDecoy reports Associated for a framework-specific but spoofable clue, Likely when independent evidence agrees, and Confirmed test when a controlled run carries valid signed provenance. Likely evidence can be a strong ARTEX-to-Norma request sequence or a verified model-API observation containing both ARTEX prompt markers and a role-coherent tool schema.
Does ARTEX attribution automatically block the request? +
No. Observed client identity, framework attribution, threat score and response remain separate decisions. The controlled request in this post was logged and allowed.
Continue your investigation
- See how WebDecoy identifies agents
Review the difference between a claimed name, verified identity and supporting evidence.
- Choose a sensor for your stack
Install edge, server or browser coverage and confirm which requests it can observe.
- Verify a crawler claim
Check one crawler name and source address against the operator's published identity data.
Share this post
Like this post? Share it with your friends!
Want to see WebDecoy in action?
Get a personalized demo from our team.