Before execution
Discover, register, scan, and quarantine capabilities before an agent can load them.
- AI discovery
- Skill + MCP scanning
- Registry admission
DefenseClaw Cisco AI security
Security governance for the entire AI agent lifecycle.
Scan skills and MCP servers before admission, inspect prompts and tool calls at runtime, pause risky actions for human approval, and export the evidence to your existing security stack.
- Connectors
- 13
- Decision modes
- Observe · Action · HITL
- Evidence rails
- 5
Guided example · Synthetic event data
Cursor attempts to send a sensitive file externally
DefenseClaw inspects a pending shell action, correlates two findings, and blocks it before execution.
Deterministic
{ "kind": "enforcement_decision", "trace_id": "trace-demo-017", "session_id": "demo-session-017", "connector": "cursor", "decision": "block", "severity": "critical", "rules": [ "secret.file-read", "shell.data-egress-pipe" ], "executed": false}{ "kind": "enforcement_decision", "trace_id": "trace-demo-017", "session_id": "demo-session-017", "connector": "cursor", "decision": "block", "severity": "critical", "rules": [ "secret.file-read", "shell.data-egress-pipe" ], "executed": false}
DecisionBlock before execution
Reason
shell.data-egress-pipe reached CRITICAL severity
Action
Write a correlated enforcement event
What DefenseClaw did — and did not do
What it did
- Inspected the event before execution
- Correlated rule evidence and applied the active severity mapping
- Recorded a synthetic enforcement event
What it did not do
- Execute the displayed command
- Send data to an external service
- Offer HITL for a CRITICAL finding
A continuous control plane
One control plane for the agent lifecycle
Admission, runtime, and evidence operate as one sequence instead of three disconnected security products.
During execution
Inspect prompts and tool calls at the connector’s strongest available interception point.
- Runtime rules
- Policy enforcement
- Human approval
After execution
Correlate each decision with durable evidence and export it to the security stack you already use.
- Audit history
- Observability
- OTLP + Splunk + webhooks
Available now
Six things you can do today
Start with one concrete control, then extend the same policy and evidence contract across the lifecycle.
Connector-aware enforcement
Use the strongest control each agent exposes
DefenseClaw normalizes the decision contract while preserving the difference between native ask, downgraded confirm, and pre-execution blocking.
| Connector | Pre-execution block | Native ask | Fail closed |
|---|---|---|---|
| Claude Code | Yes | Yes | Yes |
| Codex | Yes | No | Yes |
| OpenClaw | Yes | Yes | Yes |
| Cursor | Yes | Yes | Yes |
| Hermes | Yes | No | No |
| OpenCode | Yes | No | Yes |
| OmniGent | Yes | Yes | Yes |
Start with evidence
Put a guardrail around your first agent in five minutes.
No LLM key is required for deterministic runtime rules or static scanner checks.