Control center
Every action by every AI agent passes through one control layer.
- 47Agents found
- 18,420Actions checked today
- 18,389Passed
- 23Held for a person
- 8Blocked by policy
Today on every surface
| Surface | Events from 00:00 to 14:10 | Held | Blocked | Controls |
|---|
Needs attention
- vendor-agentThird-party agent using a long-lived keyCritical
- local-cli-agentRunning with a developer’s personal credentialCritical
- mcp-docs-gatewayNo owner and broader access than it usesHigh
Latest decisions
| Time | Decision | Agent and action | Control |
|---|
Agent fleet
Orchestrate many agents on one mission. Every branch, push and merge passes through the same security gates.
MSN-204Patch critical VULN-8841 across 3 services
Nine agents plan, patch, test, scan, review and release. A person approves the merge.
- Agents
- 9
- Pushes
- 0
- Blocked
- 0
- Merged
- 0
- Open vulnerabilities
- 3
Orchestration
Merge gates
| Pull request | Tests | Secrets | Packages | Code scan | Review | Person | State |
|---|
Live activity
Model supply chain
Secrets are removed before a prompt leaves your network.| Route | Data allowed | Retention | Decision |
|---|---|---|---|
| Self-hosted model, production-eu | Code, logs and tickets | Stays in your network | Allowed |
| Direct provider contract | Code with secrets removed | Zero retention in contract | Allowed |
| Third-party AI router | None | Reseller chain not verified | Blocked |
| 24 canary tokens | Planted in repositories and prompts | 0 used outside the company | Watching |
Agents
A live inventory of who and what is acting on your infrastructure.
| Agent | Owner | Surfaces | Credential | Mode | Risk |
|---|
Hold pauses risky actions for a person. Witness records them without blocking anything.
Control library
Each control is a policy on one shared engine, across every surface an agent can touch.
| ID | Control | Surface | Response | Severity | OWASP | Release |
|---|
Incident simulation
One of many controls, followed from the first action to the evidence.
Session ses_4f2a
deploy-agent on task PLAT-2041
Step 1 of 7
AI Operator
AI investigates and acts on incidents. A separate policy engine decides what it is allowed to do.
- AI proposesFrom an approved action list
- Policy decidesFixed rules, not AI
- Scoped accessValid for this action only
- Result verifiedRolled back if it fails
- Record signedEvery step is kept
Operator run
INC-0148: a secret was exposed during a Witness pilot.
Autonomy by action
| Action | Family | Reversible | Resources | Autonomy |
|---|
Witness explains. Suggest prepares for a person to run. Approve acts after a person agrees. Autopilot acts alone inside the guardrails.
Evidence
The same record answers security, audit and the board.
NIS2 access control
Article 21- Every agent with an owner and purpose
- Access reviews and exceptions
- Third-party agents and suppliers
NIS2 incident reporting
Article 23- Agent activity separated from human activity
- Drafts for the 24-hour and 72-hour stages
- Root cause and mitigation for the final report
DORA change management
RTS Article 17- Independent approval for each change
- Documented purpose and scope
- Fallback and emergency change records
Board oversight pack
NIS2 Article 20- Agent risk summary for management
- Decisions, exceptions and open items
- A recurring record of oversight
Warded supports your compliance work. It does not certify compliance, and your team stays responsible for every report.
One control layer for every agent
Alerts are only the output. The platform is the layer every agent action passes through.
Agents
Anything that acts
Hold points
Where actions are checked
Control plane
The shared core
Agent fleet
Many agents, one mission
AI Operator
Acts under the same policy
Modules
What teams use
Your stack
Planned integrations
What makes it infrastructure
It sits in the path
Agent actions go through it before they reach production.
It keeps the record
The action graph links people, agents, tasks and resources over time.
Others build on it
Teams and partners add policies, packs and integrations through one API.
It runs with you
Self-hosted in your environment, across any model, agent or cloud.
A control is one policy
control: KUB-01 when: surface: kubernetes verb: get resource: secrets outside_task_scope: true then: response: hold approver: on-call-security evidence: [nis2-art-23]