EXHIBIT 12 / PRODUCT ANATOMY

What is OpenClaw—and what gives it authority to act?

OpenClaw is an open-source agent product that can operate through a user’s machine, messages, and accounts. A replaceable model produces proposals; the surrounding product connects channels, tools, state, permissions, and execution.

SHORT ANSWERIt is an agent system, not a model—and its usefulness is also its permission risk.
6–9 MIN · MANUAL CONTROLS · NO ACCOUNT

ROOM 01 / DEFINE

A product name hides four distinct kinds of authority.

  1. 01 / CHANNEL Who may ask?

    Messaging channels and pairing or allowlists determine which senders can reach a configured agent surface.

  2. 02 / MODEL What may be proposed?

    A selected model interprets context and proposes text or tool calls. The model is replaceable and is not the OpenClaw product itself.

  3. 03 / TOOL What can execute?

    The active profile, allow/deny policy, sandbox, provider restrictions, and channel permissions shape which tools are visible and usable.

  4. 04 / EFFECT What changes persist?

    Commands, browser actions, messages, scheduled jobs, and account operations can create real effects, so identity, scope, approval, and audit matter.

ROOM 02 / INSPECT

Trace a request without treating the product as magic.

This teaching trace follows documented roles. It does not connect to OpenClaw, inspect an installation, or run a command.

MANUAL EVIDENCE PATHNOT A CAPTURED SYSTEM TRACE

INGRESS

A channel identifies the requester

The gateway receives a message through a configured channel and applies the channel’s pairing, allowlist, and session rules.

WHAT THIS LETS YOU SAY

Unknown senders can be stopped before their content reaches the model.

Step 1 / 4

Keyboard: focus this instrument and use ← or →. It never advances by itself.

ROOM 03 / TEST

Build a bounded permission envelope.

Change four deployment facts for a hypothetical file-writing request. The browser applies a teaching policy; it never contacts OpenClaw or a filesystem.

LIVE BROWSER CLASSIFIERNO MODEL, HARDWARE, ACCOUNT, OR REMOTE SERVICE

SCENARIOIncoming request includes untrusted email text and proposes writing a summary file.

01Requester identity

The sender is outside the paired or allowlisted set.

02Workspace authority

The run is limited to a bounded workspace and tool set.

03Untrusted content handling

A tool-disabled or read-only stage contains untrusted content first.

04Write approval

No approval record authorizes this mutation.

FIRST FAILURE / INGRESS

Reject before the message reaches a tool-enabled run.

An unknown sender is outside this hypothetical gateway’s allowed requester set.

SELECTED FACTS
  • The sender is outside the paired or allowlisted set.
  • The run is limited to a bounded workspace and tool set.
  • A tool-disabled or read-only stage contains untrusted content first.
  • No approval record authorizes this mutation.

Real deployments require current documentation, threat modelling, secret handling, network controls, updates, monitoring, and incident response beyond these four controls.

ROOM 04 / VERIFY

Read product claims and permission claims at different scopes.

OFFICIAL DESCRIPTION

An extensible open-source agent

OpenClaw describes itself as acting through a user’s machine, messages, and accounts, with replaceable harness and protocol integrations.

OFFICIAL SECURITY MODEL

One gateway, one trust boundary

Current documentation distinguishes tool policy, sandboxing, permission profiles, requester scope, and node execution controls.

NOT CLAIMED

Safety or autonomy breakthrough

The museum does not audit the codebase, certify a deployment, equate product popularity with AGI, or execute the product.

REVIEWED SOURCES

Open the record before repeating the claim.

What is OpenClaw? (Why OpenClaw) OpenClaw Foundation · key-sections · checked 2026-09-11 Project positioning, replaceable harnesses, open standards, and replaceable-infrastructure sections; every feature, release, and partner statement was not reviewed page by page. Tool and agent permissions — OpenClaw OpenClaw Foundation · key-sections · checked 2026-09-11 Control-plane tools, node execution, plugins, sandboxing, and per-agent access profiles; the full security documentation and implementation code were not reviewed page by page. openclaw/openclaw 代码仓库 GitHub · key-sections · checked 2026-09-07 Repository landing page: the README description, licence, installation method, and displayed star count; code and issues were not read.