Skip to content

Agentic attack framework coverage

WebDecoy recognizes reviewed evidence associated with agentic security frameworks while keeping client detection, framework attribution, and operator identity separate.

OBSERVE / CORRELATE / EXPLAIN

Know the framework.
Keep the uncertainty.

WebDecoy connects independent evidence across the places agentic tools leave traces—without turning a User-Agent, IP address, or shared runtime into certainty.

Target HTTPModel APIControl planeControlled test

ARTEX

Multi-plane

Correlates reviewed target clients, privacy-bounded orchestration evidence, and signed validation runs.

Target HTTPModel APIControl planeControlled test
Evidence boundary

A shared Norma client identifies the lineage, not ARTEX by itself.

AntiFox

Lineage-aware

Separates product-specific observations from the shared ARTEX-derived WebFetch lineage.

Target HTTPModel APIControl planeControlled test
Evidence boundary

Shared lineage evidence never becomes a named AntiFox attribution on its own.

PentestCode

Probe-aware

Recognizes reviewed client, model, and coherent target-probe evidence without relying on one label.

Target HTTPModel APIControl planeControlled test
Evidence boundary

Generic exploit payloads and ordinary browser traffic stay unattributed.

CyberStrike AI

Client-aware

Connects reviewed target and model clients with explicit control-plane self-identification.

Target HTTPModel APIControl planeControlled test
Evidence boundary

A copied client claim is association evidence, not proof of an operator.

PentAGI

by VXControl

Control-plane

Preserves reviewed product self-identification in the generic framework evidence contract.

Target HTTPModel APIControl planeControlled test
Evidence boundary

No target-level PentAGI attribution is claimed from generic scanner traffic.

HexStrike AI

Orchestrator-aware

Correlates a reviewed MCP orchestrator identity with privacy-bounded, multi-family tool orchestration.

Target HTTPModel APIControl planeControlled test
Evidence boundary

Nmap, Nuclei, SQLMap, or Metasploit use alone never becomes HexStrike attribution.

Cairn

Model + control plane

Recognizes a reviewed phase-prompt contract and coherent authorized control-plane inventory.

Target HTTPModel APIControl planeControlled test
Evidence boundary

Claude Code, Codex, Pi, browser, or scanner traffic alone never becomes Cairn attribution.

Strix

Model-plane

Combines an atomic model attribution-header pair with an independent root-agent orchestration contract.

Target HTTPModel APIControl planeControlled test
Evidence boundary

Ordinary Chromium, Caido, agent-browser, and scanner traffic never becomes Strix attribution.

Shannon

Model-plane

Combines a reviewed multi-stage pipeline contract with an independent bounded deliverable-tool schema.

Target HTTPModel APIControl planeControlled test
Evidence boundary

Playwright, Chromium, curl, Pi, task agents, and ordinary security tools never become Shannon attribution.

Coverage means WebDecoy has a source-backed, bounded evidence family for at least one observable plane. It does not mean every request made by that framework is distinguishable, or that anyone sending a copied string is the framework’s developer or a known operator.

Evidence plane What a compatible sensor can observe What it cannot prove alone
Target HTTP Reviewed request clients and coherent target-visible probe behavior Who operated the client, or whether an exploit succeeded
Model API Privacy-bounded fingerprints of reviewed model clients, prompt contracts, or role schemas Raw prompts, completions, tool arguments, or the target
Control plane A coherent, reviewed service shape from authorized inventory or an instrumented integration That unrelated target traffic came from the service or framework
Controlled test Signed, short-lived provenance for a property-scoped validation run That later public traffic with similar characteristics is the same framework

The registry is deliberately narrow. New or changed fingerprints enter a maintainer review path before they can support customer-facing attribution.

WebDecoy uses three attribution tiers:

  1. Associated — one reviewed but spoofable evidence family was observed.
  2. Likely — independent evidence families agree in one observation or bounded, property-scoped sequence.
  3. Confirmed test — a signed WebDecoy validation run proves that the customer initiated that controlled test.

One signal never becomes Likely merely because it has a high threat score. Shared clients remain labeled as a framework family, coarse actors cannot join a sequence, and evidence never crosses customer or property boundaries.

Read Framework attribution for the complete evidence model and Correlation detections for how bounded sequences become stable findings.

ARTEX and AntiFox share source lineage and can emit the same Norma WebFetch client. WebDecoy records that client as ARTEX lineage rather than naming either product. A product-specific observation plus independent, strongly scoped evidence can support a named framework attribution.

PentestCode coverage spans reviewed target clients, model-client evidence, control-plane self-identification, and coherent probe behavior. Generic browser requests, common exploit strings, and near-neighbor payloads stay unattributed.

CyberStrike AI coverage spans reviewed target, model, and control-plane client evidence. A client string is still self-asserted evidence: it can support association, but it does not identify the person or organization operating it.

PentAGI has reviewed control-plane infrastructure coverage. WebDecoy requires its coherent standard service shape rather than a port, title, or label alone, and reports only that an address exposed infrastructure consistent with PentAGI/VXControl. The finding stays Associated: it does not claim that a target request came from PentAGI or identify the operator. Ordinary PentAGI/1.0 SearxNG, browser, and scanner traffic remains unattributed.

HexStrike AI coverage is model-plane and orchestration-aware. A verified sensor can report only reviewed generic tool families, bounded counts, and keyed digests; raw prompts, targets, arguments, results, and credentials are neither stored nor returned. Coordinated use of multiple families creates a generic offensive toolchain association. A Likely HexStrike AI attribution additionally requires the independent, source-reviewed hexstrike-ai-mcp orchestrator identity. Common tools such as Nmap or Nuclei never name HexStrike on their own.

Cairn coverage spans authenticated model-plane observation and authorized control-plane inventory. A verified model sensor can report its source-reviewed three-phase prompt contract as a privacy-bounded digest. Inventory must match the standard port, Cairn OpenAPI identity, and project-list API shape together; a port or title alone is insufficient. Each produces an Associated finding. Ordinary Claude Code, Codex, Pi, browser, Nmap, or Nuclei activity never names Cairn.

Strix coverage is model-plane only. A verified sensor can report the exact, source-reviewed application attribution-header pair as one atomic group and the root/subagent orchestration lifecycle as a second independent group. Either group is Associated; both together may be Likely. One header alone, near-neighbor values, ordinary Chromium, Caido, agent-browser, Nmap, Nuclei, SQLMap, or payloads never name Strix.

Shannon coverage is model-plane only. An authenticated sensor can report its source-reviewed multi-stage pipeline contract and bounded deliverable-tool schema as two independent, privacy-bounded digests. Either group is Associated; both together may be Likely. Raw prompts, target URLs, credentials, tool arguments, and results are rejected. Generic Playwright, Chromium 124, curl, Pi, task-agent, browser, or security-tool traffic never names Shannon.

Open Detections, select a request, and look for Framework attribution and Correlated activity. Each finding keeps its attribution tier, evidence boundary, and links to the retained supporting detections.

You can also use WebDecoy’s hosted MCP server to search correlation findings from an AI assistant. Ask it to explain the supporting evidence and uncertainty, then follow the direct dashboard link to inspect the underlying requests.

For a direct lookup, use search_correlation_findings with finding_type: framework_attribution and a reviewed framework_subject such as pentestcode, cyberstrikeai, hexstrike_ai, cairn, strix, or shannon. Use finding_type: offensive_toolchain_association to retrieve generic multi-tool orchestration without making a framework claim. The assistant receives the public subject, tier, bounded evidence-family names or generic tool-family summary, and coverage state—not proprietary thresholds, raw prompts, or internal cohort keys.