Where does the data live?
On your servers. DurbarAI runs with Docker Compose or Kubernetes, and your documents, agent runs and audit log stay in your PostgreSQL. Model calls go to the provider you configure, Anthropic or OpenAI, with your own API key. Self-hosting details
Which controls does the product have?
- Approvals: tools you mark wait for a named approver; the decision is audited.
- Stop button: halts new runs and in-flight runs at the next step.
- Release gate: a new version is scored against the live one before people reach it.
- Audit log: per-event HMAC hash chain, Ed25519-signed checkpoints, one-click verify.
- Data-loss prevention rules on prompts, replies, tool inputs and outputs: block, redact or log, with a simulate mode.
- Document search that respects who may open each document, with citations.
- OIDC sign-on and SCIM provisioning, implemented but not yet tested against a real identity provider.
How does it map to the standards you answer to?
This crosswalk shows where DurbarAI produces evidence for a requirement. It does not make you compliant, and Prashna Labs holds no certification.
| Standard | What it asks | Evidence from DurbarAI |
|---|---|---|
| EU AI Act, Article 12 | Record-keeping: automatic logs of events over the system's life. | Hash-chained audit log with verify, search, and JSONL or CEF export. |
| EU AI Act, Article 14 | Human oversight: people can oversee, override and stop the system. | Approval-gated tools with a named approver; a stop button for new and in-flight runs. |
| NIST AI RMF, Govern | Roles, accountability and policies for AI systems. | Each agent carries an identity card with its owner, risk level and the tools that change things; role-based capabilities. |
| NIST AI RMF, Measure | AI systems are tested and evaluated before and during use. | Release gate scores each candidate against the live version on 4 quality and 2 safety metrics. |
| NIST AI RMF, Manage | Risks are acted on, including disengaging a system. | Stop button, approvals on consequential tools, and audited rollback. |
| SOC 2, CC6.1 | Logical access is restricted to authorized people. | Roles and capabilities; document search that respects who may open each document; check access explains each decision. |
| SOC 2, CC8.1 | Changes are authorized, tested and approved before release. | Every promotion passes the release gate; overrides and rollbacks are audited. Evidence bundles group events by control. |
What have we not done yet?
- Prashna Labs has no SOC 2 report and no penetration test.
- There is no multi-factor sign-in; OIDC is the one sign-on protocol.
- Audit keys are held by whoever operates the deployment, so the log is not proof against the operator.
- The stop button fails open if its own store errors.
Every known gap is on what DurbarAI cannot do yet, kept in step with the product's claims registry.
How do I report a vulnerability?
Email hello@prashnalabs.com with "Security report" in the subject. Our security.txt has the same contact. A person reads every report and replies.