01
Which agents exist, and what can they touch?
Agents show up faster than anyone registers them. Qbrin gives each one an identity and an owner, and maps the tools and credentials it can reach.
The platform
Discover. Authorize. Prove.
One evidence graph behind every agent action: who is acting, whether the action is justified, and a receipt for what happened.
Evidence graphWho is acting, and with what?
See it move
The whole platform in one short loop.
The wheel, the authorize gate, the decision engine and the rotating headline, recorded from the sections on this page.
A loop showing four screens in turn: the Discover, Authorize and Prove wheel; the authorize gate returning ALLOW, then HOLD, then ESCALATE for three example actions; the four-ring decision engine changing verdict across four example cases; and the headline “Verify before AI acts” with its last word changing.
Beyond peopleYou secured your people.
You secured your people.
What about your agents?
Three questions every team running AI agents ends up asking, usually after something has already happened.
02
What do they inherit from whoever launched them?
An agent that borrows a person’s login inherits everything that login can do. Qbrin scopes what an agent may do to the action in front of it.
03
Was the action actually justified?
Logs tell you what happened after the fact. Qbrin checks the claim against current evidence before the tool runs, and holds what the evidence does not support.
Inside the platform
Three moves between an agent and your tools.
01 · Discover
Every agent gets an identity. Nothing runs anonymously.
Register the agents, tools and credentials in your stack and see what each one is allowed to reach, before it takes its first action.
- An identity and an owner on every agent
- Every tool and credential mapped to the agent that holds it
- A reasoning trace for every run
Agent inventoryIllustrative
AgentOwnerToolsLast actionStatus
refund-agentfinance-ops4refund_paymentgoverned
support-triagesupport6tag_ticketgoverned
data-sync-botplatform3export_rowsheld
research-agentgrowth5web_searchgoverned
onboarding-agentpeople-ops2create_accountnew
02 · Authorize
Verify before AI acts.
Every proposed action is checked against current evidence, permissions and company policy. Decisions come back in low milliseconds as ALLOW, HOLD or ESCALATE.
- Evidence freshness and bindings verified
- Prompt injection filtered before a tool runs
- An expiring signed grant; tools fail closed without it
POST /v1/authorizeIllustrative
refund_payment₹4,200 · order #8841
- Order found in billing system
- Amount within what was paid
- Inside the refund policy
ALLOWgrant · qg_9f2c…e71a · expires in 60s
03 · Prove
Every decision leaves a receipt.
Approved actions carry a signed grant. Every decision, allowed, held or escalated, lands in a tamper-evident audit trail you can replay.
- Tamper-evident audit log of every decision
- Human sign-off on high-risk steps
- Replay any run from its trace
Audit trailIllustrative
Chain verified4 entries · 0 gaps
- 10:41:07refund_payment ₹4,200ALLOW#a91f…3c2e
- 10:41:33refund_payment ₹95,000HOLD#c40b…77d1
- 10:42:10delete_customer_recordsESCALATE#5e08…0aa9
- 10:42:58send_email → customerALLOW#2b7d…91f4
The control plane
Everything between an agent and your tools, in one place.
Discover
Agent identity & inventory
An identity, an owner and a scoped credential for every agent, with each tool it can reach mapped in one place.
Agent governanceAuthorizeEvidence-backed authorization
POST /v1/authorize checks each action against current evidence, permissions and policy, then answers ALLOW, HOLD or ESCALATE.
Authorization APIEnforceTool boundary policy
Policy is applied before a tool runs. Downstream tools and executors fail closed unless a valid grant is presented.
Read the docsContainSandbox containment
Agents work inside a sandbox they cannot escape, so a wrong decision stays small.
How containment worksProveTamper-evident audit
Every decision lands in an audit trail you can replay, with human sign-off on high-risk steps.
Security modelPlatform tour
Follow one action from registration to receipt.
A four-step walk-through of what happens when an agent tries to refund a customer.
Step 1 of 4 · Register the agentIllustrative
refund-agentowner · finance-ops
governedCredential
scoped key ••••8f2aTools
Sandbox
onTrace
every run recordedThe decision engine
Four checks, one verdict, no guessing.
Before a tool runs, the gate asks four questions. An action only gets a grant when every ring is full. Anything short of that is held or sent to a person.
- Evidencefresh, bound to the claimOrder and invoice match the claim
- Authoritywho is askingRequester verified
- Policytool boundaryInside the refund policy
- Injection screeninstructions in the dataNothing injected
ALLOW
Illustrative. Fill shows how much of each requirement the action meets.
The trust layer, in motion
One verification layer for every domain — from your sources, through qbrin, to your agent
Your agent asks. qbrin pulls the evidence from your live systems and knowledge, verifies every claim against it, and only then lets the answer through — the same way whether it’s a security alert, an outage, a mission decision, or a payment. Verified decisions, or an honest “not enough evidence,” never a confident guess.
- Cybersecurity
- IT operations
- Defence
- Space & aerospace
- Finance
- Healthcare
- Legal
- Manufacturing
0 / 500fabrications on nonexistent-entity trapsLlamaIndex: 11 · naive RAG: 155+
93.7%precision when answeringLlamaIndex: 84.7% · naive: 80.5%
88%recall@20 on the retrieval benchbge-m3 dense
74.9%answer coverage — abstains, never guessesheld-out, audited benchmark
Measured on a 1,000-question HotpotQA A/B against a tuned RAG baseline — the honest floor, not a cherry-pick. See the full benchmarks ↓
Verify before AI acts
One gate at the tool boundary. Evidence, permission and policy, checked before anything runs.
Qbrin Labs
Measured, not claimed.
Every number here comes from a committed benchmark, with its sample size and its caveats. Where we do not lead, the write-ups say so.
0invented answers across 120 trap questions
declined, or corrected with the cited real fact
0%Citations you can trustWith its double-checking layer on, a cited source genuinely backs the answer 86% of the time, raw keyword search manages just 6%. (Measured, ERB.)
0%Always the current factWhen a fact changes, qbrin returns the new value every time; mem0 (open-source) keeps the stale one 3 times in 4. (Measured, n=8.)
0%Finds the right sourceRetrieval lands the correct source document in the candidate set 88% of the time, and the double-checking layer then keeps the cited ones honest. (Measured, ERB.)
Engineering notesAll posts
- A safety gate that blocks 100% of attacks and 100% of real work looks perfect on the scoreboardWhy substring grounding checks reject valid operator commands.
- The most dangerous prompt injection is stopped by retrieval, not by detectionCredential-scoped boundaries at the retrieval plane.
- A 400-token cap on our verifier was silently dropping correct answersHow a fixed limit turned parser errors into silent refusals.
See it answer
See it answer your hardest question.
Bring one real question your team keeps re-asking. We'll connect a source, read-only, and show you the answer, sourced, in seconds, in a 20-minute walkthrough. Nothing changes in your tools.
20-minute walkthrough
- One source connected, read-only
- Your real question answered, with sources
- Nothing changes in your tools