AI agent governance AI agent control infrastructure

Control what your AI agents can do.

Before data leaves. Before money moves. Before systems change.SANSAR checks protected agent actions against your policies before execution. Allow approved work, require human approval for sensitive actions and block requests that break your rules.

Request a demo → See the product

One checkpoint · allow · hold · deny

01 / Agent requestsEvery protected action
02 / SANSAR checksIdentity · policy · region · approval
03 / AllowProceeds and is recorded
HoldWaits for the named approver
DenyBlocked before execution
The console

See it. Stop it. Prove it.

See which agents are acting, what they requested and why an action was allowed, held or denied. Intervene from one command centre for the agents SANSAR governs.

SANSAR · Command centreProduct illustration
1

Every agent, by name.

Each governed agent with its owner, policy and live activity.

2

Every action, on its lane.

Requests flow through policy, sovereignty and evidence gates in real time.

3

Every verdict, counted.

Allowed, escalated, denied and halted, with sovereignty enforcement state.

4

One switch to stop.

A global halt stops execution across every agent SANSAR governs. Try it above.

Product illustration drawn from the SANSAR command centre: same agents, gates and verdicts, presentation data. Try the global halt.
DeterministicExplicit rules decide. An AI model never makes the policy call.
Fail-closedIf a required condition is not met, the action does not run.
VerifiableSigned, tamper-evident records you can check yourself.
Data boundariesDestination and region rules checked before data moves.
YoursSelf-hosted, sidecar or in your own cloud. Your keys.
Illustrative decisions
The problem

Access to a tool is not
permission for every action.

01 / Data & AI teams

Keep transfers within policy.

Set approved data destinations. Block protected transfers that fall outside your policy.

02 / Platform & operations

Control changes to live systems.

Decide which actions run automatically and which wait for approval.

03 / Governance & assurance

Review the decision and evidence.

Inspect the policy, the approvers and the signed record behind each action.

01 / The decisionIllustrative demo · no real data

One request.
Three possible answers.

Choose a workflow. The same checkpoint reaches a different verdict, for reasons anyone can read.

The agent asks

SANSAR checks
    The result

    Illustrative policy example. No transfer, payment or change is performed.Inside the console →
    02 / Inside the consoleMonitor · enforce · approve · prove

    Your agent asks.
    Your rules decide.

    Enforce policy at the point of action. Deterministic checks allow permitted requests, hold those requiring approval and deny those that fail. Decisions follow explicit rules applied to the request and relevant state.

    01

    See who is acting.

    Every governed agent in one fleet view, with the policy it runs under and the person who owns it.

    Monitor
    02

    Enforce the boundary.

    All six policy rules passed, and the sovereignty gate still refused the transfer. Every decision shows which gate decided, and why.

    Policy & sovereignty
    03

    Route the exception.

    The named approver sees the amount, counterparty, data class and precedent, against a clock. If time runs out, the request is denied.

    Human oversight
    04

    Trace the decision.

    Inspect the request, policy decision and approval in a linked evidence chain. Verify signed records to detect changes.

    Cryptographic evidence
    How the verdict is reached
    An AI model does not make the policy decision.
    Agent fleetPresentation data
    Agent fleet: all 36 registered agents online, with type, policy and owner for each. Decision record DEC-900000: a 12,000 euro transfer denied by the sovereignty gate after all six policy rules passed. Escalation dossier: a Tier 1 critical request pending human review, with SLA timer, amount, counterparty, data class, precedent, and approve or deny controls. On expiry it auto-denies. Evidence chain verified: 14,400 entries intact, each linked to the previous entry hash.
    01 / 04Close-ups from the product · seeded presentation data
    03 / DeploymentYour environment · your keys

    Runs where
    your data lives.

    The checkpoint belongs close to the systems it protects. Choose the shape that fits your security model; we confirm supported paths with your team during evaluation.

    01

    Self-hosted

    Inside your network, on infrastructure you operate.

    02

    Sidecar

    Beside each agent runtime, so checks happen next to the action.

    03

    Your cloud

    Deployed in your own cloud account, under your identity and keys.

    04

    Managed

    Operated by SANSAR, for teams that want to move first.

    Integrate

    Connect in-process through the SDK, or place the MCP gateway in front of your tools.

    SDK → MCP gateway →
    04 / EvidenceCryptographic proof & attestation

    Check the record. Detect changes.

    Inspect a signed record of the request, policy decision and required approvals. Verify its integrity and compare execution evidence with the connected system’s records.

    The browser example runs a real signature and tamper test on synthetic data. A valid signature proves a record is unchanged; attestation scope is defined per integration.

    Run the verification yourself →
    Decision recordIllustrative · unsigned
    R-00184
    01
    Who requested it?
    sales-agent-07
    02
    What was requested?
    file.export → supplier / US
    03
    Which rule applied?
    eu-only-destinations.v3
    04
    Could approval override it?
    No exception permitted by this rule
    05
    What decision was recorded?
    DENY / destination not permitted
    Request · decision · destination evidence01 / 01
    05 / Why nowIncident evidence · regulation & guidance

    AI agents act.
    You remain accountable.

    An authorised agent can still make an unauthorised payment, disclose sensitive data or change the wrong system. Organisations need enforceable limits, human oversight and a traceable record.

    What went wrong

    Jul 2025

    An agent deleted production data.

    Replit confirmed its agent deleted data from a user's app database. It was restored; development and production are now separated by default.

    ↗     CVE-2025-32711

    An email became a route to data.

    A published case study describes a zero-click prompt injection in Microsoft 365 Copilot. It was patched.

    ↗     Nov 2025

    Agents were used to execute attacks.

    Anthropic reported a campaign against roughly 30 targets, succeeding in a small number of cases.

    ↗

    What the guidance calls for

    EU AI Act

    Traceability and human oversight.

    Can you trace the action? Can a person intervene?European Commission ↗
    NIST AI RMF

    Govern, map, measure, manage.

    Who owns the risk? How is the control evaluated?NIST AI Resource Center ↗
    OWASP

    Complete mediation. Least privilege.

    Check downstream requests against policy. Require approval for high-impact actions.Excessive agency guidance ↗

    Incidents illustrate failure modes; they are not SANSAR customer stories or tests of its effectiveness. Framework applicability depends on the system, use case, role and jurisdiction, and does not establish certification or compliance.

    06 / QuestionsTechnical due diligence

    Before your next deployment.

    Does SANSAR use AI to decide whether an action is allowed?+

    No. SANSAR uses deterministic policy checks: explicit rules applied to the request and the relevant state, including permissions and approvals. An AI model does not judge whether another agent may act.

    Does every agent action pass through SANSAR?+

    Coverage is defined by the deployment. Complete mediation requires every route to a protected action to be checked, including calls through intermediary services. We review network paths, credentials and downstream access with you during evaluation.

    Can we stop all agents at once?+

    Yes, for agents under SANSAR's control. The global halt stops governed agent execution. Each integration defines how in-flight actions are handled; stopping does not reverse completed actions.

    What does cryptographic proof establish?+

    Signature verification establishes that a record is unchanged and associated with a signing key. Attestation identifies the stages and evidence it covers. A valid signature alone does not prove an action was safe or that a downstream operation succeeded.

    How does SANSAR support data sovereignty?+

    Set rules for approved destinations and regions, then check agent requests against them on connected routes. Processing locations, storage, model providers and subprocessors still need review across the wider system.

    Request a demo

    Put your agent workflow
    to the test.

    1. 01Define what the agent may do and where its data may go.
    2. 02See which actions proceed, wait for approval or get blocked.
    3. 03Review control coverage, failure behaviour and the evidence.

    Request your demo

    We use your details to respond to your enquiry. Privacy notice

    info@sansarai.ai
    SANSAR · Real console · unedited
    Recorded from the running product at real speed. Seeded presentation data, isolated from production.