Research note 001 22 Nov 2026

Autonomous pentesting needs evidence, not just autonomy.

ARTEX Research examines the operating loop behind AI security agents: define authorized scope, choose a useful action, preserve proof, and stop for review when judgment matters.

  • Authorized scope
  • Traceable proof
  • Human review
research_run_001
interface study
Objective

Prove an authorization boundary can be crossed.

in review
01
Scope accepted

Authorized target and testing window are attached.

api.example.com · owner-approved
02
Hypothesis planned

A tenant identifier may bypass object ownership checks.

risk: horizontal access · no destructive payload
03
Proof captured

Request, response, command, and affected asset stay linked.

traffic / api-get-order · status 200 · tenant mismatch
04
Review human

Impact is validated before the finding becomes a report.

approval required · evidence version pinned

Static interface study. The production ARTEX workspace lives at artex.lol.

The question

What makes an autonomous security result trustworthy?

A tool call is not a finding. A fluent explanation is not evidence. The useful unit is a bounded run: a target, an approved scope, a sequence of decisions, the traffic or command output behind each claim, and a human who can review the result.

ARTEX is organized around that unit. Its public console keeps tasks, assets, exploration, traffic, findings, commands, and approvals connected so an operator can inspect how a conclusion was reached.

This research page is independent from the application itself. It describes the model of work; the console is the place where that work runs.

Scope defines permission. Actions produce records. Evidence supports claims. Review assigns consequence.

Operating loop

Move from intent to inspectable action.

The loop stays deliberately narrow. Each step creates a record that the next step can use and a reviewer can challenge.

  1. 01

    Frame the scope

    Record the system that may be tested, the rules of engagement, the available credentials, and the actions that require approval.

    output / authorization record
  2. 02

    Plan the next test

    Turn observations into a concrete hypothesis. Choose the least invasive action that can confirm or reject it.

    output / decision + rationale
  3. 03

    Execute and record

    Invoke the tool, capture the request and response, preserve timestamps, and connect the artifact to the asset under test.

    output / command + traffic + asset
  4. 04

    Review the consequence

    Separate technical reproduction from business impact. A person decides whether the evidence is sufficient and what happens next.

    output / reviewed finding

Proof standard

A finding must survive inspection.

The report begins with the artifact that supports it. Every claim should point back to the observable behavior that made it useful.

Claim only ?
“The endpoint appears to allow cross-tenant access.”
  • No request or response attached
  • No affected identity or asset
  • No reproduction path
Inspectable proof ✓
Tenant A reads tenant B’s order after changing one object identifier.
  • Request and response preserved
  • Identity, asset, and record linked
  • Impact reviewed and scoped
Reproduction

Can another reviewer reach the same observation?

Keep the exact command, request, response, and dependency context available.

Attribution

Does the evidence belong to the asset in the finding?

A matching domain or timestamp is a filter. Ownership and scope determine whether the artifact is relevant.

Impact

What can an authorized actor actually do with it?

Technical behavior becomes a security finding only after its consequence is stated and reviewed.

Freshness

Does the proof still describe the current system?

Retests should refer to the original finding and the new artifact, not merely repeat the original narrative.

Operating constraints

Autonomy expands only inside explicit limits.

Permission is not inferred from technical reach. Scope, approval rules, and the consequences of an action remain explicit parts of the run.

AUTH

Use only systems you own or are authorized to test.

Written permission and a defined testing window come before the first request.

STOP

Escalate sensitive actions to a human operator.

Destructive, high-impact, or ambiguous actions require a review point rather than an optimistic default.

SCOPE

Keep targets, tools, and evidence in the same record.

Asset inventories and blocklists are controls. A guardrail matters only when the workflow actually checks it.

REVIEW

Treat generated confidence as a claim, not proof.

Humans decide impact, remediation priority, and whether a result is ready for a customer report.

Continue into the console

Read the evidence. Then run the workflow.

ARTEX combines a planner, parallel security agents, traffic capture, assets, findings, and human approval in one inspectable workspace.

Open ARTEX