Security incident response platform with evidence collection, action execution, and audit capabilities. Provides read-only evidence collection tools for AgentTeams Workers and controlled response execution with approval workflows.
CyberGuard exhibits significant definition quality gaps. Of 15 tools, 9 (alert.snapshot, intel.lookup, network.search, boundary.policy, endpoint.timeline, asset.context, recovery.metrics, evidence.validate, knowledge.search) have generic, under-20-character descriptions that fail to explain WHEN to use them or WHAT they return. The 6 action tools (block_ioc, disable_account, isolate_endpoint, quarantine_workload, terminate_process, disable_persistence) have better descriptions but lack output schema documentation. Parameter descriptions are present but often generic ('Tool-specific arguments' is not actionable). No tool documents its return structure. Error handling guidance is absent across all tools. Naming is verb-noun compliant but lacks clarity on tool distinctions (e.g., 'alert.snapshot' vs 'endpoint.timeline', when does an agent choose one over the other?). The 'arguments' parameter in discovery tools is a catch-all that violates single-responsibility principle.
Retrieve alert snapshot for incident investigation
Retrieve asset context for incident investigation
Block indicator of compromise (IOC) - L1 risk, reversible action
Retrieve boundary policy information for incident investigation
Disable user account - L2 risk, reversible action
Disable persistence mechanism - L2 risk, irreversible action (requires host_lab mode)
Discovery tools (alert.snapshot, intel.lookup, network.search, boundary.policy, endpoint.timeline, asset.context, recovery.metrics, evidence.validate) have descriptions under 20 characters, violating the 10 - 1024 character baseline. Descriptions like 'Retrieve alert snapshot for incident investigation' lack actionable context: what fields are returned? When should this be called vs endpoint.timeline? What does 'incident investigation' entail?
No tool documents its output schema. LLMs cannot plan downstream calls or extract required fields. For example, does alert.snapshot return incident_id, severity, timestamp, affected_assets? Without this, agents cannot chain tools effectively.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Retrieve endpoint timeline for incident investigation
Validate evidence for incident investigation
Perform intelligence lookup for incident investigation
Isolate endpoint from network - L2 risk, reversible action
Search knowledge base for incident investigation
Search network evidence for incident investigation
Quarantine workload - L2 risk, reversible action
Retrieve recovery metrics for incident investigation
Terminate process - L2 risk, irreversible action (requires host_lab mode)
Discovery tools accept a generic 'arguments' parameter (type: object, description: 'Tool-specific arguments') that violates single-responsibility and parameter clarity principles. This forces LLMs to guess what sub-fields are valid. Replace with explicit, typed parameters (e.g., filter_type, time_range, severity_level).
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. E.g., if alert.snapshot fails because incident_id is invalid, what should the agent do? Search for incidents first? Retry? This violates the recovery-guide pattern.
Action tools (block_ioc, disable_account, isolate_endpoint, quarantine_workload, terminate_process, disable_persistence) lack dry-run or confirmation mechanisms. Irreversible operations like terminate_process should support a confirmation step before execution, per the confirmation-request pattern.