Skip to content

Framework attribution

A request tells you which client reached your site. Framework attribution asks the harder question: which larger automation framework launched that client?

WebDecoy keeps those answers separate. A detection can say “Observed client: Norma WebFetch” while independently reporting “Likely ARTEX,” with the evidence and confidence tier shown beside it.

Autonomous security and scraping frameworks rarely use one unique HTTP signature from start to finish. They delegate work to libraries, enrichment services, browsers, and command-line clients. Blocking or labeling on one User-Agent therefore produces two bad outcomes:

  • a shared library is mislabeled as every framework that happens to use it; or
  • the individual requests are detected, but the campaign behind them remains invisible.

Framework attribution connects the sequence without erasing what WebDecoy directly observed.

Tier What it means Example
Associated One framework-specific but spoofable clue was observed artex-enrich/*, ARTEX’s enrichment client User-Agent
Likely Independent evidence families agree within one observation or a bounded sequence A verified model-API sensor sees ARTEX’s prompt contract and role-coherent tool schema together; or ARTEX enrichment is followed by Norma WebFetch under the same strong identity
Confirmed test An authenticated WebDecoy user initiated a property-scoped controlled run, and the request carried valid, expiring signed provenance A WebDecoy-labeled ARTEX validation run

“Confirmed test” confirms the controlled test run. It does not claim that arbitrary public traffic with the same User-Agent came from that framework.

ARTEX is the first framework supported by this attribution model.

ARTEX can use Norma WebFetch to retrieve a target. Norma is also a standalone client that other software can run, so User-Agent: norma/0.4 is evidence of Norma and not proof of ARTEX.

WebDecoy upgrades the framework attribution only when it sees independent evidence:

  1. ARTEX’s framework-specific enrichment client and Norma WebFetch are observed on the same property within ten minutes; and
  2. the requests share a strong, tenant-scoped identity, or the same address plus a matching cryptographic, device, or JA4 request fingerprint.

The supporting detection is linked from the detection detail view, so you can inspect the evidence instead of accepting an opaque label.

A model-API sensor can see orchestration evidence that the target HTTP server cannot: the agent’s system-prompt contract and the tool schemas offered to the model. WebDecoy accepts a privacy-bounded summary of those signals through its authenticated SDK route.

The artex.model-plane.v2 rule assigns Likely ARTEX only when one verified observation contains both:

  1. the two stable ARTEX prompt markers in the rule’s registry; and
  2. a coherent role schema: at least four registered tools for worker or planner, or all three registered setup tools.

If only the prompt contract or only the coherent role schema is present, the observation is retained as Associated ARTEX. This is intentionally weaker: one evidence family is useful investigation context, not likely attribution. Valid observations with neither family are ledgered as evaluated and do not produce attribution, so reconciliation does not reconsider them forever.

The summary contains indicator IDs and SHA-256 fingerprints. It cannot contain raw prompts, completions, tool arguments, request bodies, or target names. Unknown fields and indicators are rejected.

Runtime and transport clues do not qualify on their own. In particular, omp/*, a GLM model name, curl, an address, or a model endpoint does not establish ARTEX. Oh My Pi is a general-purpose runtime that other software can use.

The v2 migration preserves existing model-plane findings, clears only the v1 scan ledger, and lets the bounded reconciler re-evaluate historical observations once. That backfill corrects setup and partial-evidence coverage without duplicating v1 and v2 findings.

An authenticated model-API sensor may add run_correlation_id to the bounded framework observation. The reporter must compute it as lowercase HMAC-SHA-256—or an equivalent 256-bit keyed digest—of its local run or session identifier. It must never send the raw identifier. WebDecoy rejects arbitrary labels, short IDs, uppercase encodings, and run digests without a sensor namespace.

The artex.instrumented-run.v1 rule joins observations only when all of these values agree:

  • organization and property from the authenticated SDK key;
  • sensor name and observation schema version;
  • the 64-character keyed run digest; and
  • a ten-minute observation window.

Two coherent role schemas produce Associated ARTEX for the instrumented run. The rule upgrades the run to Likely ARTEX only when coherent setup, planner, and worker schemas are all present and the ARTEX prompt contract is also observed. The finding links the supporting detections and records the roles, window, sensor, schema version, and digest; it does not change the request score or create an enforcement rule.

Missing run IDs remain standalone model-plane observations. WebDecoy never falls back to IP address, User-Agent, model, or runtime to join them. The same digest in another organization, property, sensor, schema version, or outside the window is a different namespace and cannot join. Inside one namespace and window, digest equality is the reporter’s assertion that observations belong to one run: reporters must use a stable per-run source identifier and a sensor-held key so accidental reuse does not coalesce separate runs. A run with more than 100 retained observations is marked coverage-limited and cannot be promoted to Likely by this rule.

The keyed digest follows the detection and correlation-evidence retention policy. Because the source identifier and HMAC key never leave the reporter, WebDecoy cannot reverse the digest into the local run ID.

WebDecoy deliberately refuses several shortcuts:

  • Norma alone: identifies the client library, not the framework that launched it.
  • A shared or coarse actor fingerprint: groups unrelated clients and cannot establish that two requests came from one party.
  • A copied framework header: User-Agents and ordinary request headers are client-controlled.
  • A high threat score: scoring estimates request risk; it is not framework identity.
  • A matching IP alone: NAT, proxies, and shared infrastructure make an address insufficient.
  • A runtime or model alone: shared runtimes and hosted models identify infrastructure, not the framework using it.

These limits reduce false attribution and prevent a crowded fingerprint from inheriting one member’s behavior.

For a controlled framework validation, WebDecoy issues a short-lived token bound to one organization, property, framework, and test run. The edge sensor relays the opaque token, and ingest verifies its signature, expiry, and exact tenant scope before storing decoded provenance.

The bearer token itself is discarded. A token is valid for at most 15 minutes, survives service-key rotation through a receive-only previous key, and cannot be used to raise the threat score or trigger enforcement.

Open a detection to see the framework attribution directly beneath the evidence summary:

  • the framework name;
  • Associated, Likely, or Confirmed test;
  • a plain-language explanation;
  • the directly observed client; and
  • a link to any supporting detection.

Framework attribution is an investigation aid. Threat scoring and response actions remain separate decisions.

Correlation is scoped to your organization and, for framework sequences, the property where the requests were observed. WebDecoy does not join one customer’s traffic to another customer’s actors.

Model-plane evidence can corroborate the same framework family seen from a target, but WebDecoy does not claim that two separately collected events share an operator or run unless a valid tenant-scoped correlation key links them.

  • View detections: inspect request evidence and supporting detections
  • Actors: understand persistent client identity across IP rotation
  • Verified agents: cryptographic and network proof for published crawler identities
  • Detection coverage: choose a sensor that can see raw HTTP clients