Products AI 360

AI 360 · AI agent security & governance

Every AI agent, model and token. Governed automatically.

AI agents don't type. AI 360 finds every agent in your estate, routes each prompt to the right model, blocks policy violations before they happen, and shows what every agent costs.

See the live dashboard
  • Read-only discovery
  • Runs in your region
  • Fails closed
Blackbead Tenant Demo bank 2 CI
◌ Demo mode — this tenant runs on a synthetic AI estate.
Overview
3 things need you today
1 critical posture failure · 2 tool calls waiting for approval
SecurityFinanceAI platform
AI assets
12
3 found this week
Invocations · 30d
15.5k
all via the proxy
Posture checks
22 / asset
OWASP LLM · NIST AI RMF
Business cases
6
ROI ◌ estimated
Highest-risk agentsSorted by exploitability
AgentPlatformRiskFailing checksStatus
claims-triage-agentCopilot StudioCritical[n] / 22Exploitable
kyc-doc-summariserMCP serverHigh[n] / 22Latent
hr-policy-copilotCopilot StudioHigh[n] / 22Awaiting approval
fx-research-genieDatabricks GenieMedium[n] / 22Latent
Waiting for approvalTool calls
hr-policy-copilot → files.delete
Destructive tool · needs a second approver
claims-triage-agent → crm.update
Write tool · needs approval
Routed spend vs baseline◌ illustrative
AI 360 Overview, Security view. Synthetic demo estate — no customer data.
22
posture checks per AI asset
5
checks on every MCP tool call
9
routable models in the catalogue
0
write calls during discovery

Your security stack was built for humans. Your AI estate isn't human.

Agents you didn't approve are already in production — calling APIs, reading sensitive data and spending budget. Tools built for people logging in miss them. Pick your seat at the table:

For the CISO

Find the agents nobody approved — before an attacker does.

Run a read-only discovery
  • Read-only discovery. Copilot Studio, Databricks Genie and your MCP servers, with credentials from your vault.
  • 22 posture checks per asset, mapped to OWASP LLM Top 10 and NIST AI RMF.
  • PII and prompt-injection scanning on every call that passes the proxy.
  • Exposure ranked by exploitability, with data from OSV.dev and CISA KEV.

Platform

Two planes. One call path.

Agents call AI 360 instead of a vendor. Every call is governed and metered on the way through.

Callers
Agents & copilots
Copilot Studio, Databricks Genie, in-house agents
AI 360runs in your region · fails closed
Governance planeFind, assess and gate every agent and tool call
Economics planeRoute, meter and prove the value of every call
Models
9 routable, cloud or on-box
Tools
Your MCP servers

Governance plane — know it, assess it, gate it.

See everything
Every agent, model and MCP server, discovered read-only.
Know the risk
22 checks per asset against OWASP LLM Top 10 and NIST AI RMF.
Gate risky actions
Five checks before any tool call runs.
Track exposure
Exposure from OSV.dev and CISA KEV.

Economics plane — route it, meter it, prove it.

The right model
The right model per prompt, with the reason recorded.
Know the cost
Spend per user, model and business case.
Hold the service level
Latency and availability per agent.
Start governed
New agents start governed, behind the two-person rule.

AI model proxy

The right model for every prompt. And the reason why.

Hard constraints like data residency apply first; only then are the remaining models scored on quality and cost. Regulated data never leaves its region.

  • 9 routable models, cloud and on-box
  • Residency and policy filter before scoring
  • Every decision logged with its reason
Routing decisions
kyc-doc-summariser
→ on-box Qwen · residency
fx-research-genie
→ [model] · quality
hr-policy-copilot
→ [model] · cost ceiling
claims-triage-agent
→ [model] · quality
Routing decisionExample
Agent
kyc-doc-summariser
Data class
PII detected
Constraint
Residency — stay in region
1 · Hard constraints
[model]excluded · leaves region
Other cloud modelsexcluded · leaves region
2 · Scoring
Selected on-box Qwen✓ in region

Reason: regulated data — residency removed cloud models before scoring.

Model proxy · decision log. Example decision on synthetic data.

MCP gateway

Agents reach tools through MCP. Every call passes five checks first.

New tools are denied by default. If the gateway can't decide, the call doesn't run.

  1. 1
    AllowlistOnly approved tools. New ones denied by default.
  2. 2
    Argument scanPII and injection checked in every argument.
  3. 3
    Human approvalDestructive tools wait for a person.
  4. 4
    Fail closedPlane unavailable? The call is blocked.
  5. 5
    Execute & logRuns, and leaves an audit record.
Pending approvals · 2 of 2
hr-policy-copilot → files.delete
Destructive tool · two-person rule applies
High
✓ Passed
Allowlist
✓ Clean
Arguments
◷ 1 of 2
Approval
Armed
Fail closed
Not run
Execute
Arguments
{ "path": "/policies/archive/2023/",
  "recursive": true }
Requested by
hr-policy-copilot (agent)
First approval
SOC analyst · approved
Needs
A second, different approver
Decision note (required)
Why are you approving or denying this call?
Nothing runs until a second person approves.
MCP gateway · approval for a destructive tool call. Synthetic example.

How it works

From first connector to a governed estate in four steps.

Step 1

Discover

Connect Copilot Studio, Databricks Genie and your MCP servers with read-only credentials from your vault.

Step 2

Assess

Every asset gets 22 posture checks and an exposure scan, ranked by what an attacker can actually use.

Step 3

Govern

Agents call the proxy instead of a vendor. Policy, residency and cost ceilings apply on every call.

Step 4

Measure

Spend, SLOs and ROI per business case, ready for the risk committee and the CFO.

Why Blackbead

Built for regulated estates. Honest about every number.

Read-only by design

Discovery makes zero write calls to your estate.

Fails closed

When a check can't run, the call doesn't either.

Residency first

Region rules apply before any model is scored.

Explainable routing

Every decision carries its reason.

Never claims a call it didn't make

Routed and executed calls are labelled apart.

Humans stay in the loop

Destructive actions and publishing need a person.

Get a free AI estate assessment.

30 minutes with the founder and one read-only connector. You leave knowing which agents you have, what they cost and where they're exposed.

Book the assessment or email info@blackbead.ai