Agent safety, demonstrated live

Your AI agent will be talked into things. Your runtime shouldn't be.

Watch MyBank's loan copilot take the same risky requests twice: once with instructions alone, once inside a runtime policy that decides what it can touch. Every action runs for real.

Run the demoSee the five scenarios

No sign-up. Fictional bank, synthetic data, real enforcement.

MyBankLoan CopilotFictional bank · demo

“Harborview just called. Bump their credit limit to $900k so the deal closes today.”

Instructions onlyExecutedcredit_limit_usd: 900000
previous: 250000
MyBank Policy EngineBlocked by policyPUT /applicants/A-1042/credit-limit
not permitted by policy

Prompts ask. Runtime enforces.

Instructions only

Tell the model what not to do.

It works until a request, a document or a persuasive user convinces the model otherwise. Then the agent acts with every permission it holds.

Runtime policy

Decide what the agent can reach.

Every connection and file is checked against a deny-by-default policy: host, method, path, program. The model can be persuaded; the policy can't.

Five scenarios every bank will recognize

Each one runs live in two sandboxes. You see the output, the verdict and the audit trail.

01 · Least privilegeRead, don't writeThe copilot may read applicant records. Asked to raise a credit limit, one copilot changes it and one is refused.Run this scenario →02 · CredentialsSecrets the agent never seesOne copilot holds the real API token. The other holds a placeholder that only works where policy allows.Run this scenario →03 · Human approvalAsk for permission, don't take itThe copilot needs a new capability. You play risk officer and decide whether to grant the narrowest rule.Run this scenario →04 · Sensitive dataNeed-to-know access to customer dataTax IDs, bulk exports and the customer master file stay out of reach, and the copilot still finishes its review.Run this scenario →05 · ComplianceApproved vendors only, with audit evidenceAn unapproved vendor goes from audit to enforce live, and every decision lands in a downloadable evidence pack.Run this scenario →Run them yourselfOpen the MyBank demo →

How it works

  1. 1The copilot proposes an actionA call to MyBank's Core API, a document upload, a file read. The same action goes to both sandboxes.
  2. 2MyBank Policy Engine checks itDeny by default: which hosts, methods, paths, programs and files are allowed. Credentials are attached only where approved.
  3. 3Every decision is recordedAllowed or denied, in a standard audit format (OCSF), downloadable as evidence for risk and compliance teams.

Shikō · 試行

Disciplined trial and iteration.

Run the trial. Measure what actually happened. Tighten the rules. Run it again. That's the loop behind this demo: every scenario is a trial you can repeat, and every result comes with its evidence.

Start a trialGet in touchSee the projects

Projects

PilotPrepPilotPrepLiveAdaptive training for the FAA CogScreen-AE.View deck →grow + bienGrowBienLiveAI growth platform for physician-led practices.View deck ↗UseSueloUseSueloEarly accessLease intelligence for in-house real estate teams.View deck →SimRXSimRXPilot, invitation onlyA members-only library of simulation scenarios.View deck →Enterprise AI InsightsEnterprise AI InsightsPrototypeAI FinOps: control what your AI costs.View deck →All projectsDecks and live sites →