Radware Agentic AI Protection for OpenClaw

OpenClaw integration package for Radware Agentic AI Protection.

Use this package when OpenClaw is already installed and running. The package does not install or onboard OpenClaw.

What This Package Provides

• Out-of-path enforcement plugin: registers OpenClaw lifecycle hooks and calls Radware before tool execution.
• In-path setup helper: adds a Radware OpenAI-compatible model provider to an existing OpenClaw config.
• Production-safe config merge: updates the existing OpenClaw config, writes a backup, and stores no secrets in openclaw.json.
• Portal-controlled policy: Block and Report blocks, Report Only reports and allows.
• Local fail behavior: fail-close or fail-open for Radware API availability failures in out-of-path mode.

Before You Start

Run the commands as the same OS user that owns the OpenClaw config and runs the OpenClaw gateway.

Confirm the config exists and is already onboarded:

test -f "${OPENCLAW_HOME:-$HOME}/.openclaw/openclaw.json" && echo "OpenClaw config found"

If OpenClaw uses a custom config path, pass it with --config /path/to/openclaw.json.

Set runtime secrets through your shell, systemd EnvironmentFile, Kubernetes Secret, or secret manager. Do not write real keys into openclaw.json.

export RADWARE_INPATH_API_KEY="sk-rdwr-..."
export RADWARE_INPATH_BASE_URL="https://api.agentic.radwarecto.com/v1/openai"

export RADWARE_OUT_OF_PATH_API_KEY="sk-rdwr-..."
export RADWARE_OUT_OF_PATH_URL="https://api.agentic.radwarecto.com/llmp/digester/agentic-api"

export LLM_MODEL="gpt-4o"
export RADWARE_USER_IDENTIFIER="openclaw-out-of-path"
export RADWARE_FAIL_MODE="fail-close"

Out-of-path deployments also require the customer's normal LLM provider credentials, such as OPENAI_API_KEY, because Radware does not replace the LLM provider in out-of-path mode.

Upgrade From 0.1.0

Version 0.1.1 and later are production-safe by default:

• The setup helper expects an existing onboarded OpenClaw config.
• The setup helper refuses --in-path --out-of-path and also refuses adding one Radware path to a config that already contains the other Radware path.
• The docs present in-path and out-of-path as mutually exclusive deployment options.

If you previously ran the 0.1.0 setup helper on a fresh server before OpenClaw onboarding, it may have created a partial openclaw.json. If OpenClaw reports existing config is missing gateway.mode, move that partial file aside and onboard OpenClaw first:

mv ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bad-partial
openclaw onboard --mode local

Then rerun setup with the latest package and choose exactly one path:

npx -y -p openclaw-radware-agentic-protection@latest radware-openclaw-setup --in-path --set-default-model

or:

openclaw plugins install openclaw-radware-agentic-protection@latest
npx -y -p openclaw-radware-agentic-protection@latest radware-openclaw-setup --out-of-path

Package Version Policy

Use openclaw-radware-agentic-protection@latest for customer deployments. Version 0.1.4 and later include the production-safe deployment flow, foreground gateway guidance, and strict one-path-per-OpenClaw-deployment guardrails.

Do not unpublish previous NPM versions as a normal maintenance action. Unpublishing can break pinned installs and those version numbers cannot be reused. If a previous version should no longer be used, deprecate it with an upgrade message instead:

npm deprecate openclaw-radware-agentic-protection@"<0.1.4" \
  "Please upgrade to 0.1.4 or later. Earlier versions were superseded by production-safe OpenClaw deployment guidance and stricter in-path/out-of-path validation."

Deprecation preserves reproducibility for pinned users while showing a clear install-time warning.

Choose One Integration Path

Choose exactly one integration path for a given OpenClaw deployment. Do not configure in-path and out-of-path together in the same OpenClaw deployment. To evaluate the two approaches, use separate OpenClaw environments or separate change windows.

Option A: In-Path Only

Use in-path when Radware should sit inline on OpenClaw provider traffic.

Dry-run the change:

npx -y -p openclaw-radware-agentic-protection@latest radware-openclaw-setup \
  --in-path \
  --set-default-model \
  --dry-run

Apply the change:

npx -y -p openclaw-radware-agentic-protection@latest radware-openclaw-setup \
  --in-path \
  --set-default-model

This adds a provider named radware-openai that uses:

• apiKey: ${RADWARE_INPATH_API_KEY}
• baseUrl: ${RADWARE_INPATH_BASE_URL}
• model: ${LLM_MODEL} or gpt-4o

The customer OpenAI API key is not used by this in-path provider entry.

Restart the OpenClaw gateway with the environment variables loaded.

openclaw gateway --force is a foreground server process. Seeing gateway] ready means the gateway started successfully; the command is expected to keep running until stopped. Run it in the customer's normal service manager, terminal multiplexer, or a dedicated shell.

Option B: Out-Of-Path Only

Use out-of-path when the customer wants to keep the existing OpenClaw LLM provider and have Radware check tool actions before execution.

Install the plugin:

openclaw plugins install openclaw-radware-agentic-protection@latest

Dry-run the config merge:

npx -y -p openclaw-radware-agentic-protection@latest radware-openclaw-setup \
  --out-of-path \
  --dry-run

Apply the config merge:

npx -y -p openclaw-radware-agentic-protection@latest radware-openclaw-setup \
  --out-of-path

This adds:

{
  "plugins": {
    "entries": {
      "radware-agentic": {
        "enabled": true,
        "hooks": {
          "allowConversationAccess": true
        },
        "config": {
          "apiKeyEnv": "RADWARE_OUT_OF_PATH_API_KEY",
          "endpoint": "${RADWARE_OUT_OF_PATH_URL}",
          "model": "${LLM_MODEL}",
          "userIdentifier": "openclaw-out-of-path",
          "failMode": "fail-close",
          "enforcementMode": "portal-decision"
        }
      }
    }
  }
}

Restart the OpenClaw gateway with the environment variables loaded.

openclaw gateway --force is a foreground server process. Seeing gateway] ready means the gateway started successfully; the command is expected to keep running until stopped. Run it in the customer's normal service manager, terminal multiplexer, or a dedicated shell.

Inspect the plugin:

openclaw plugins inspect radware-agentic --runtime --json

Manual Change-Control Installation

For production environments that require manual review:

1. Run the relevant --dry-run command.
2. Review the generated JSON.
3. Merge the generated provider or plugin entry into the existing openclaw.json.
4. Keep all secrets in environment variables.
5. Restart OpenClaw through the customer's normal process.

Portal Policy Semantics

enforcementMode: "portal-decision" means the Radware portal controls the decision:

• Block and Report: Radware returns IsBlocked: true; the plugin blocks the tool call.
• Report Only: Radware returns IsBlocked: false; the plugin allows the tool call and the portal records the event.

failMode applies only when Radware is unavailable or returns an invalid response:

• fail-close: block/pause the tool action if Radware cannot be reached.
• fail-open: allow the tool action if Radware cannot be reached.

Policy blocks are never bypassed by fail-open.

Non-Destructive Validation

Validate from an OpenClaw staging channel or a low-risk test agent.

Minimum checks:

• Benign prompt: should be allowed.
• AI Guardrails credit-card PII test: should follow the portal policy.
• HAPBlocker test: should follow the portal policy.
• Blocked medical/medicine topic: should follow the portal policy.
• Behavioral tool/action test: use an existing low-risk tool, such as a test email or dry-run write action. Out-of-path enforcement only runs when OpenClaw is about to execute a tool.

Recommended flow:

1. Start the OpenClaw gateway and wait for gateway] ready.
2. Send a benign prompt through the customer's normal OpenClaw channel or Control UI.
3. Send the AI Guardrails prompts through the same channel and verify Radware portal events.
4. For out-of-path Behavioral validation, trigger a real low-risk tool action. A chat-only prompt will not call the plugin.
5. For in-path Behavioral validation, make sure the provider request includes tool/action context. A normal chat prompt may only validate AI Guardrails.

For each event, record:

• OpenClaw integration path: in-path or out-of-path.
• Radware portal User Name.
• Model.
• Expected result.
• Actual result.
• Event ID.
• Module: AI Guardrails or Behavioral / Agentic Protection.

Troubleshooting

Setup Fails With Missing Config

Run the helper as the same user that runs OpenClaw, or pass the exact config path:

npx -p openclaw-radware-agentic-protection radware-openclaw-setup \
  --out-of-path \
  --config /path/to/openclaw.json \
  --dry-run

Setup Fails With Missing gateway.mode

The helper expects an already onboarded OpenClaw config. This usually means the wrong OS user or config path was used, or OpenClaw was not onboarded yet.

For a lab-only fresh install, onboard OpenClaw first, then rerun the helper.

Out-Of-Path Does Not Generate Events

Out-of-path enforcement runs before tool execution. If the OpenClaw workflow only chats with the LLM and does not call a tool, the plugin has nothing to enforce.

Gateway Looks Stuck After Start

openclaw gateway --force starts the gateway in the foreground. It is healthy when the log shows gateway] ready; leave it running in that terminal, run it under the customer's service manager, or stop it with Ctrl+C.

In-Path Does Not Show Expected Behavioral Result

For Behavioral validation, the provider request must include tool/action context. A normal chat prompt without a tool/action context may only exercise AI Guardrails.