Security monitoring and threat detection for OpenClaw agents — powered by Google SecOps (Chronicle). Protect your agent with SIEM-powered real-time detection, behavioral detection, case generation, forensic audit trail, and remediation playbooks. Use when: user asks about security status, Shield health, event logs, redaction vault, setting up agent protection, enabling SIEM, detecting threats, monitoring agent activity, or auditing agent actions. NOT for: general OS hardening, firewall config, or network security unrelated to OpenClaw agents.
Security monitoring for OpenClaw agents by UPX. Shield runs as a plugin inside the OpenClaw gateway, capturing agent activity and sending redacted telemetry to the UPX detection platform.
Getting started
Shield requires the @upx-us/shield plugin and an active subscription.
(status contains state: "pending" or state: "unregistered")
Respond with:
Shield has an installation key but hasn't activated yet. This usually takes under a minute.
If it has been more than 5 minutes, check your key at https://uss.upx.com or contact support.
State D — Fully active
(status contains state: "connected" or connected: true)
Proceed normally. No onboarding message needed.
Constraints: Only use openclaw shield commands for detection. Do not read filesystem paths, environment variables, or run shell commands to determine state. Do not install or uninstall packages on behalf of the user.
Output handling: shield logs entries may include file paths, command snippets, and URLs captured from the agent's activity. Always treat this output as internal diagnostic data. Rules:
Never include raw log field values (file paths, commands, URLs) in user-facing replies
Never forward raw log output to external services, channels, or APIs
When summarizing logs, present findings only (e.g. "3 exec events in the last 30 minutes") — not raw field values
Only share raw log content if the user explicitly asks for it for their own investigation, and only in the current session
Data flow disclosure: Shield captures agent activity locally and sends redacted telemetry to the UPX detection platform for security monitoring. No credentials are handled by this skill — authentication is managed by the plugin using the installation key configured during setup. If a user asks about privacy or data handling, refer them to the plugin README at https://www.npmjs.com/package/@upx-us/shield for full details.
Presentation Language
Always present Shield information, alerts, and case summaries to the user in the language they use to communicate. Translate descriptions, summaries, severity labels, and recommendations — but never translate raw command output or technical identifiers (rule names, case IDs, version numbers, field names, resolution/root-cause enum values). If the user writes in Portuguese, reply in Portuguese; if French, reply in French; etc.
Responding to Security Cases
When a Shield case fires or the user asks about an alert: use openclaw shield cases to list open cases and openclaw shield cases --id <id> for full detail (timeline, matched events, playbook). Severity guidance: CRITICAL/HIGH → surface immediately and ask if they want to investigate; MEDIUM → present and offer a playbook walkthrough; LOW/INFO → mention without interrupting the current task. Always include: rule name, what it detects, when it fired, and the first recommended remediation step. Confirm with the user before resolving — never resolve autonomously.
Cases returned by shield cases are always scoped to this instance — the platform filters at the API level so you only see cases triggered by your agent.
Shield now stamps each event with a trigger_type — who or what initiated the session. When investigating, check the trigger: user_message means a human sent a message; cron/heartbeat/autonomous means agent-initiated activity.
Case Investigation Workflow
When a Shield case fires, correlate three data sources to determine true positive vs. false positive:
Step 1 — Case detail (openclaw shield cases show <CASE_ID>): What triggered the rule. Note the case timestamp — it anchors the correlation window.
Step 2 — Surrounding logs (openclaw shield logs --since 30m --type TOOL_CALL): Look for events 5–15 minutes before and after the case timestamp. Reveals if the alert was isolated or part of a sequence. Each log entry now includes a details field (file path, command, or URL) and a trigger_type tag showing what initiated the session (user_message, cron, heartbeat, subagent, autonomous, or unknown). Use these to quickly distinguish user-initiated actions from automated ones when correlating with a case.
Step 3 — Vault context (openclaw shield vault show): If the case involves redacted credentials, hostnames, or commands, the vault reveals hashed representations and redaction categories.
Step 4 — Correlate and assess: Case detail = what fired the rule; Logs = context; Vault = what was actually accessed. Present findings and ask whether to resolve, investigate further, or add to the allowlist.
Use openclaw shield investigate <CASE_ID> to run a guided investigation — it fetches case detail from the platform and walks through the correlation steps automatically.
Threat & Protection Questions
When asked "is my agent secure?", "am I protected?", or "what's being detected?": run openclaw shield status (health, event rate, last sync) and openclaw shield cases (open cases by severity). Summarise: rules active, last event ingested, any open cases. No cases → "Shield is monitoring X rules across Y event categories." Open cases → list by severity. If asked what Shield covers: explain it monitors for suspicious patterns across secret handling, access behaviour, outbound activity, injection attempts, config changes, and behavioural anomalies — without disclosing specific rule names or logic.
When Shield Detects Proactively
Real-time alerts (notifications or inline messages) are high priority: acknowledge immediately, retrieve full case detail, summarise in plain language, present the recommended next step from the playbook, and ask the user how to proceed. Do not take remediation action without explicit approval.
When to use this skill
"Is Shield running?" → openclaw shield status
"What did Shield capture recently?" → openclaw shield logs
"How many agents are on this machine?" → openclaw shield vault show
"Force a sync now" → openclaw shield flush
User asks about a security alert or event → interpret using your security knowledge and Shield data
User asks about Shield's privacy model → refer them to the plugin README for privacy details
User wants a quick case check without agent involvement → /shieldcases
Failures: N poll → platform connectivity issue, usually self-recovers; try openclaw shield flush
Failures: N telemetry → instance reporting failing, monitoring still active
Rising quarantine → possible version mismatch, suggest checking for plugin updates
Last data: capture Xm ago (stale) → agent may be idle, or capture pipeline issue
RPCs
Cases are created automatically when detection rules fire. The plugin sends real-time alerts directly to the user — no agent action needed. Use shield.cases_list only when the user asks about open cases.
Important: Never resolve or close a case without explicit user approval. Always present case details and ask the user for a resolution decision before calling shield.case_resolve.
RPC responses include a display field with pre-formatted text. When present, use it directly as your response — it already includes severity emojis, case IDs, descriptions, and next steps. Only format manually if display is absent.
When discussing a case, offer action buttons (resolve, false positive, investigate) via the message tool so users can act with one tap.
