Security
Security and data handling
Where your code goes, what leaves your network, and what we can prove. This page states the posture that exists today in present tense, and labels everything that does not exist yet as roadmap.
True today
The posture today.
- scanning
- Scanning runs against your repositories and produces structural facts: services, APIs, events, tables, ownership. The output is the graph and its evidence, not a mirror of your codebase.
- evidence + ledger
- Evidence and the hash-chained ledger live in Findry's Postgres instance deployed for your pilot. One pilot, one database.
- model calls
- Model calls are key-gated and off by default. Deterministic fallbacks are the default: extraction, impact analysis, and the governed pipeline run without a model key. LLM output is never evidence.
- simulation
- Simulated execution is always labelled. A simulated stage carries its simulated flag verbatim in the attestation; it is never omitted.
- sign-in
- Sign-in federates with your identity provider via OIDC (Okta, Azure AD, Google). Accounts are invite-only: a first login never self-provisions an account, and disabling a member revokes their sessions immediately.
- roles
- Three roles with approval separation enforced at the gate: an approver cannot approve their own change, and the refusal is recorded.
- secrets
- Integration credentials are envelope-encrypted at rest (GCP KMS, or a labelled local key), masked in every API response, and writes fail closed when no encryption provider is configured.
- verification gates
- Every agent change passes five sealed gates before rollout: static analysis, secret scanning over the changed files, test execution, an independent review, and a structural-quality check. Reviewer identity is sealed alongside the writer's, and gates fail closed on any error. The reviewer's findings are advisory and recorded in full for the approver: a language model's judgment is never the thing that passes or fails a change.
- governance caps
- Org-level allowlists for agents, models, and deploy adapters, with hard caps on blast radius, changed files, and tokens per run. A cap that cannot be verified refuses the run, and every refusal is sealed with the control named.
- export
- The evidence ledger is exportable, and an offline verifier re-checks the hash chain with no access to Findry. Audit events stream as JSONL for your SIEM, and any Blueprint exports an EU AI Act Article 12 audit pack.
- third parties
- No code is retained by third parties through Findry.
On the roadmap
- saml bridge
- SAML 2.0 federation for identity providers that cannot speak OIDC. OIDC federation is live today.
- scim
- Directory-group provisioning and deprovisioning. Invite-based membership with immediate session revocation is live today.
- soc 2 program
- A formal SOC 2 program. We do not claim SOC 2 compliance today, and will not before a report exists.
Security questions from your review team are welcome: [email protected]. We answer with what is true, in the same tense it is true in.