Governance, Threat Intelligence and Validation

AgentGuard combines local protection with centralized governance. Security teams can define policies, handle approvals, inspect events, and produce evidence organized by environment, host, or agent.

Policy management#

The policy editor provides Observe, Balanced, and Strict modes. Administrators can configure allow, alert, approval, or block outcomes for five risk categories:

• Destructive commands.
• Remote code execution.
• Data exfiltration.
• Credential access.
• Deployment.

Policies also maintain protected paths, blocked domains, blocked commands, and allowed-command lists. They can be bound to environments, hosts, or agents, distributed with versioning, and cached locally after retrieval.

Events and approvals#

Protection Events brings risk events, guardrail matches, and approval records into a central view. Teams can filter by environment, host, or agent and replay a session timeline.

The approval inbox supports centralized review of actions waiting for a human decision. This complements the one-time authorization mechanism described in Runtime Protection.

Environment security reports#

Reports can be scoped to an environment, host, or agent. They summarize assets, coverage, runtime events, guardrail matches, policy-hit rankings, supply-chain scans, and threat intelligence.

PDF export supports management reporting and audit preparation.

Threat intelligence and local self-checks#

AgentGuard supports signed threat advisories covering skills, plugins, MCP servers, supply-chain dependencies, malicious URLs, and prompt injection. Organizations can also publish their own internal advisories.

Host self-checks compare advisories with installed skills and plugins, MCP configuration, npm and pip dependencies, and DSH inventories. A match generates response recommendations such as isolation, removal, URL blocking, or token revocation, and is reported back to management.

Scheduling is supported through native tasks in OpenClaw, QClaw, Hermes, and DSH, as well as system cron and Windows Task Scheduler. Private deployments can distribute intelligence locally, including in air-gapped environments.

Patrol and checkup#

Daily patrol#

Daily patrol checks skill integrity, credential exposure, network exposure, scheduled tasks, recent file changes, attack patterns in audit logs, environment configuration, and trust-registry health.

Security checkup#

Checkup evaluates code security, trust hygiene, runtime defense, credential protection, Web3 protection, and configuration posture, producing a visual report.

These workflows help teams identify changes and maintenance needs beyond a single blocked action.

Validate policies and agents#

AgentGuard includes AI red-team probes for prompt injection, jailbreaks, tool misuse, excessive agency, and memory poisoning. Probes map to OWASP risk categories and MITRE ATLAS techniques and can retest runtime policies and online agents with scorecard outputs.

These built-in probes are part of AgentGuard's validation workflow. Axora Arena is a separate product for scoped adversarial testing and authorized assessment.

Framework alignment and audit use#

Policy versions, approval records, event timelines, and reports provide traceable evidence for internal reviews and external audits. Component capabilities and management roles support least-privilege decisions.

AgentGuard's personal-data detection categories are organized with reference to sensitive personal information under China's Personal Information Protection Law. OWASP and MITRE ATLAS mappings help teams relate checks to existing threat models.

A staged adoption approach#

This is a rollout model, not a command-level deployment guide. For deployment boundaries and identity options, see Deployment, Privacy and Integrations.