IRONHIDEData rights
Data rights are not a price axis and not a feature you toggle. Plans are
metered on protected agents and execution volume
(plans & pricing); there is no privacy tier
to buy, no opt-in flag to set, and no per-agent opt-out. Every run is
owned — the model is identical on the free Developer tier and on every
paid tier, and effective_data_rights is always "owned".
What Ironhide keeps — the environment, not your trajectory#
When your agent runs an episode, the thing Ironhide keeps and reuses is the environment your agent ran against — the mock world, its state, and the attack it faced — not your agent's own trajectory or prompts. That environment is shadowed and scrubbed, then added to a growing shared attack-episode library (it started at roughly 150 finance, CRM, and other workflow environments). Harder environments make the next customer's tests sharper, and that shared library is what keeps the free tier free.
Your agent's own run — its trajectory, its tool calls, and the verdict it
earned — is yours. You can always see and export your own results in
full; see Exporting results. The stamp is
owned, everywhere, for everyone.
The rails that still hold#
"Every run is owned" is not "anything goes." Two mechanical exclusions still gate what can ever enter a shared or sellable dataset, and neither is a data-rights choice:
| Rail | What it excludes |
|---|---|
Preview / scaffold exclusion (is_sellable) | Observed-state (Layer 2/3) verdicts are stamped maturity: "preview", and any preview — or seeded / scaffold — row is machine-excluded from every export, aggregate, benchmark, and mining path, line by line. Preview output is never sold, benchmarked, or trained on. |
| Held-out redaction | A held-out episode's spec — its labels, planted canaries, and world tripwires — is its answer key. Evidence that could quote a planted canary is masked in every customer projection, so an answer key never leaks through a verdict or an export. |
Exports carry a further protection: at export time your agent id is replaced with a random-salted, per-export pseudonym, so exported data cannot be joined back to your registration, and two separate exports cannot be joined to each other.
Your runs still power your own CI gate#
The pre-deploy CI gate answers one question: did your agent regress on your own runs since your last baseline? To answer it honestly over a stochastic agent, the gate computes a statistic over your agent's own recent runs.
- Same-tenant only. The gate reads only runs recorded by your agent, computes a statistic for you, and shows it only to you.
- None of the shared doors. The gate never exports, pools across customers, benchmarks, mines, or sells anything.
Gating your own deploys on your own runs is not export, benchmark, or resale — it is you using your own results, which is always yours to do.
Customer-Isolated Spec Vaults#
On the Business tier and up, custom mock environments and the runs recorded against them live in a Spec Vault: a per-tenant, customer-isolated store.
- Your vault's environments and runs are visible only to your tenant.
- Runs in a vault feed only that tenant's own CI gate — they are never aggregated across customers and never enter a shared or sellable dataset.
- The same preview / scaffold exclusion that keeps a preview verdict out of a sellable export keeps vault contents inside the vault.