AI agents your security team can approve
Every agent runs as an attested workload identity on your own cluster, in your cloud, on your metal, or air-gapped. Every task it does is cited and signed, so your security officer can verify what ran, where, as whom, and from what knowledge.
The problem
Today's AI agents act on stale memorized knowledge, cannot show what they read, and leave no record a security review can verify. So they stall at the security review, or they run without one.
What we build
Each status below is honest as of 2026-10-24. Available means proven on our lab clusters as of 2026-10-24, not deployed for customers. In development is being built; Planned is designed, not built. We have no customers yet and hold no certification or authorization.
Attested identity
Status: Available. Each agent is a SPIFFE workload identity issued by SPIRE on its own cluster (shown in the lab for one seeded agent). Secrets come from OpenBao to that identity alone.
Signed, cited work
Status: Available. Each task record is an in-toto statement signed with cosign and verifiable offline (shown on a synthetic example). A public transparency log is planned.
A hardened cluster per customer
Status: Available. Talos Linux with Secure Boot and TPM-sealed disks, no inbound port, destroyed and recreated from code, reconciled by Flux. Proven on lab hardware with a software TPM.
Your security officer's choices, as data
Status: In development. An assurance profile per cluster: key custody, node attestation, sign-in strength, audit retention, agent policy. Impossible combinations are rejected before anything is built.
Sandboxed agent runtime
Status: Planned. One agent per sandbox (gVisor or Kata), admission policy by Kyverno, task permissions by OpenFGA.
Your cloud, your metal, or air-gapped
Status: Planned. Customer clusters on major clouds, bare metal and disconnected sites. Today only our lab runs.
Learn more
Who it is for
Defense and air-gapped sites
With CMMC, impact-level and disconnected requirements.
Next step
The agent catalog uses O*NET data under CC BY 4.0. Task text in the catalog comes from the O*NET 31.0 Database by USDOL/ETA, used under CC BY 4.0; not reviewed or endorsed by USDOL/ETA.