Agents your security team can sign off on
Each line starts with what changes for your security, risk and platform teams, then says how DurbarAI does it. Where something is not finished, the same line says so.
Pass the security review
The questions that end most agent pilots, answered before the review meeting.
Your documents and audit log stay on your own servers
DurbarAI runs on your infrastructure with Docker Compose or Kubernetes. Documents, runs and the audit log live in your PostgreSQL.
Today: You build the container images from source.
Model calls go only to a provider you already approved
Bring your own Anthropic or OpenAI key. The provider and model are set per deployment, so moving to another approved provider is a configuration change.
People sign in through the identity provider you already run, and leavers lose access
OIDC single sign-on handles sign-in, and SCIM provisioning creates and removes accounts from your directory.
Today: Not yet tested against a real identity provider, and there is no multi-factor sign-in.
Weak settings cannot reach production
With the environment set to production, DurbarAI requires sign-in, rejects placeholder secrets and keys shorter than 32 characters, and requires https public URLs.
Keep a person in charge of risky actions
Agents read and draft freely. The steps that change something outside the conversation wait for a person.
Refunds, deletions and customer messages wait for a named person to say yes
Mark the tools that change things as approval-gated. The run stops before the call and waits for an approver; every other tool runs without asking.
Today: Approvals cover tool calls, not what an agent writes in a chat reply.
Approvers decide with the whole request in front of them
The approvals inbox shows the tool, its arguments, the agent and the policy that required the approval. The decision is recorded under the approver’s identity.
One action halts every agent when something goes wrong
An admin types a reason and halts agents. New runs are refused, running ones stop at their next step, and users see the reason.
Today: If the switch’s own store errors, runs continue. It does not cover document intake.
Nobody can push a risky change alone
Publishing an agent, changing release-gate settings and overriding a blocked release are separate permissions, and an override needs a second approver.
Answer the auditor without a scramble
When someone asks what an agent did and who allowed it, the answer is a search away.
You can show what an agent did last Tuesday, and who approved it
Agent events, approvals and policy changes go into a hash-chained audit log. Search it, open an event, and export what you find.
Today: Some run receipts are written best-effort, so a missing one is not flagged on its own.
Anyone with access can check that the log has not been edited
Verify recomputes the whole chain and points at the first event that does not match. Signed checkpoints fix the state of the log at a moment in time.
Today: The keys belong to whoever operates the deployment, so the log exposes edits by anyone without them, not by the operator.
Policy changes cannot slip by unrecorded
Policy changes are written strictly: if the audit write fails, the change fails too.
Your auditor gets signed evidence instead of screenshots
Evidence bundles group events by SOC 2 control and by week, and each bundle is signed. Reviewers verify a bundle and attest access reviews on the same screen.
Today: This is evidence your own audit can use. Prashna Labs has no SOC 2 report.
Your security team sees agent activity in the SIEM it already watches
Filter audit events and export them as JSONL or CEF.
Keep sensitive data where it belongs
Agents work with your documents and your customers’ details, so access rules and data rules travel with every run.
A restricted document never reaches someone who may not open it
Agents search only the documents the person asking is allowed to read.
Today: Search matches words, not meaning. Retrieval by meaning is on the roadmap.
Users and reviewers can check every answer against its source
Answers cite the passages they were built from.
Card numbers and personal data get blocked or redacted before they travel
Data-loss prevention rules run on prompts, replies, tool inputs and tool outputs. Each rule blocks, redacts or only logs.
New rules go live without breaking the workflows people rely on
Data-loss rules and tool policies can run in simulate mode before you enforce them.
People see why access was refused, rule by rule
Check access explains why a person can or cannot do something, so a refusal does not turn into a support ticket.
Stop agents getting worse after a change
An agent that worked last week can slip after a prompt or tool change. Each update has to hold up before it ships.
An update that scores worse than the live version does not ship
Every new version is compared with the live one on four quality and two safety metrics, whether it is published from the Studio or promoted through the API. A blocked version leaves the live one in place.
Today: The gate scores how the agent works through a task, not whether the final answer is right.
Engineers know what to fix when a release is blocked
A blocked result names the metric that failed.
Gate results hold up when someone asks how they were scored
The scorers are deterministic, so the same run always gets the same score. An AI-graded check can block a release only after it has been calibrated against human labels.
Every score points back to exactly what was tested
Eval Studio pins each experiment’s dataset and evaluators in a hash-pinned manifest.
A bad release can be undone, and the undo is on the record
Rolling an agent back to an earlier version is written to the audit log.
The people who approve actions also see quality and cost slipping
Drift and budget alerts sit next to the approvals inbox and blocked promotions in Policy & Trust.
Get agents into real work sooner
The controls above are built in, so teams spend their time on the agent and its tools.
Teams build and test agents without touching live systems
Agent Studio covers instructions, tools, knowledge and skills. The preview never runs gated tools, and people only ever reach the published version.
Every agent has a named owner and a stated risk level
Each agent carries an identity card with its owner, its risk level and the tools that change things, which gives you an inventory of the AI systems you run.
Your own product calls agents over a plain REST API
Run and stream agents from your code, with streaming over server-sent events. People can also reach agents in the web app, a widget on your site, or Slack.
Today: Slack and the widget have not yet been shown against a live customer workspace.
Agent runs show up in the monitoring you already use
Runs emit OpenTelemetry GenAI spans and JSON logs to the collector you point them at.
Every known gap in one place: what DurbarAI cannot do yet.
Get one email when DurbarAI goes open source
We write once, on the day the repository opens. No newsletter. Ask, and we delete your address.