Warded
The control plane for AI agents on live infrastructure.
Design partner brief. The demo is a simulation with fictional data.
AI agents now act on production. Most teams can’t prove what they did.
of organizations cannot clearly tell agent activity from human activity.
of large firms in the UK, France and Germany can regularly produce an AI audit trail.
What the audit log shows
user: system:serviceaccount: platform:deploy-agent
It names the credential. It does not name the person, task or cause.
Sources: Cloud Security Alliance, November 2026; Datadog and Censuswide, October 2026.
The AI supply chain is now part of your attack surface.
prompts carried planted secrets in a test of AI providers and resellers.
AI providers were found compromised, with attackers siphoning prompts.
What happened next
Stolen bait keys were used tens of thousands of times in AWS API calls. Some privacy flags were not honored.
Source: Ars Technica on research by HD Moore, November 2026.
One control layer between every agent and production.
Agents
AI IDEs, CLI agents, MCP clients, CI/CD and vendor agents.
Hold points
kubectl, Terraform, SSH, Git, cloud CLIs, SQL, MCP and model routes.
Control plane
Agent identity, action graph, policy engine, approvals and a signed ledger.
Outcome
Each action is allowed, held for a person or blocked, with evidence.
Designed to work with the agents and tools you already use.
Run many agents on one mission, with a gate on every merge.
Triage
Find affected services.
DonePatch
Three agents, three repos.
Secret blockedVerify
Test, scan and review.
PassedMerge
A person approves.
Needs youRelease
Canary, then rollout.
QueuedTests
Secrets
Packages
Code scan
Review
Person
AI can act. It can never approve itself.
| Level | What the AI does | Example |
|---|---|---|
| Witness | Explains what it would do | Shows the incident chain |
| Suggest | Prepares an action for a person | Proposes a stricter policy |
| Approve | Acts after a person agrees | Rotates an exposed secret |
| Autopilot | Acts alone inside guardrails | Quarantines an agent |
| Never | Refused at every setting | Deleting data, changing IAM, disabling a control |
Built for regulated European teams.
Runs with you
Designed to be self-hosted, with secrets masked before anything is stored or sent to a model.
NIS2 evidence
Agent activity separated from human activity, with drafts for the 24-hour and 72-hour stages.
DORA change records
Independent approval, documented purpose, and fallback records for each change.
Warded is designed to support compliance work. It does not certify compliance.
Tell us what you think. Join us from the start.
Your honest opinion
Is this a real problem for your team, and what is missing? A short reply is enough. No call needed.
Founding partners
If the idea fits, get involved from the start and help decide what Warded becomes.
Early investment
If you believe in it, we are open to discussing early investment.
Warded is a project by Rastezh Digital EOOD. This brief is not a formal offer of securities.