See every agent. Control every action. Prove every outcome.
One control plane for agent identity, permissions, runtime enforcement, human approval, incident response and audit evidence—built for autonomous software that can reach enterprise data and take real actions.
Additive infrastructure—not a rip-and-replace of IAM, SIEM, CASB or your model platform.
Agent Trust
Production agent estate
Identity, authority and live decisions
12
Agents
3
Need review
1
Blocked now
Latest decisions
- Allowed
Read customer records
Revenue research
- Approval
Issue $18,400 refund
Support resolution
- Blocked
Send data to unknown tool
Document analyst
Identity
Know every agent
A single enterprise inventory of agents, agentic applications, MCP servers, tools, service identities, credentials and owners — each bound to a durable identity that survives a model or provider change.
Control
Control every action
Runtime decisions at the moment of action: allow, deny, redact, rate-limit, sandbox, require approval or terminate — evaluated against delegated authority and current context.
Evidence
Prove every outcome
Structured, exportable records sufficient for security operations, audit, compliance, investigation and cost governance — from one event stream, without duplicate integration work.
The problem
Agents blur several legacy security boundaries at once
The product exists because autonomous agents blur several legacy security boundaries at once. Agents can inherit user context, hold service credentials, call tools, delegate to other agents, consume untrusted content and execute high-impact actions. No single legacy category is designed to represent the full decision chain from intent through model reasoning to tool execution.
Failure modes we map
- Unknown agents
- Excessive permissions
- Orphaned credentials
- Unsafe delegation
- Prompt injection
- Data exfiltration
- Runaway spend
- Weak attribution
No single legacy category is designed to represent the full decision chain from intent, through model reasoning, to tool execution.
Where we draw the boundary
- Focus the proprietary layer on agent identity, authority graphs, runtime action decisions and replayable evidence.
- Borrow existing enterprise signals through integrations rather than rebuilding every adjacent control.
- Never become a generic endpoint, network or human IAM suite.
BOUNDARY
Agent Trust Cloud is designed to work alongside existing security systems rather than duplicate them.
How it works
Every consequential action passes a decision point
By design, the cloud control plane manages inventory, policy and analytics, and gateways near your workloads enforce sensitive actions.
Reference flow
Agent
Any framework or provider
Agent Trust Gateway
Intercept
Identity + policy + risk
Decide
Protected tool / API
Execute
Event + evidence
Record
Agent
Any framework or provider
Agent Trust Gateway
Intercept
Identity + policy + risk
Decide
Protected tool / API
Execute
Event + evidence
Record
Enforcement path
Authenticate
Classify
Policy
Approve / deny
Execute
Record outcome
Authenticate
Classify
Policy
Approve / deny
Execute
Record outcome
Priority use cases
Connect a real agent, see its authority, apply a policy, watch the result
In the interactive demo you pick an agent, see its effective authority, apply a policy and watch the decision it produces.
Discover shadow agents and orphaned service identities in cloud, source code and AI platforms.
Block unauthorized data exports or dangerous tool calls at runtime.
Require human approval for high-impact actions such as payments, deletion, privilege changes or bulk external messaging.
Record every agent/tool transaction with actor, delegation chain, policy result and evidence.
Attribute model/tool consumption to agent, team, application and business purpose.
MVP DEMO
Agent attempts a risky action → Agent Trust Cloud explains the risk → policy blocks or requests approval → evidence is recorded.
Monitor
Observe the policy decision without blocking.
Approve
Route high-impact actions to a human with full context.
Block
Stop the action deterministically and record why.
The same policy runs in all three modes, so teams can prove a control before they enforce it.
What you are buying
Security outcomes buyers can inspect before they choose a plan
Start with the job your team needs done. Every outcome links to the working surface, product module or proof behind it—then pricing shows exactly which plan turns it on.
Find the agents nobody registered
Discover agents, MCP servers, service identities and credentials before asking teams to maintain another inventory.
Inspect the product →
02See what each agent can actually reach
Calculate effective authority across identity, credentials, tools, delegation and downstream systems.
Inspect the product →
03Stop unsafe actions at runtime
Allow, deny, redact, sandbox, rate-limit or require approval in the path of the action itself.
Inspect the product →
04Give auditors the decision trail
Produce evidence from the same event that made and enforced the decision, not a document assembled later.
Inspect the product →
Published plans
Choose the level of control, not a mystery package
Start here
Discover
Find every agent and understand its authority. No enforcement, no card, no time limit.
$0 / month
Annual commitment rate
- Protected actions
- Monitor only
- Environments
- 1
- Evidence retention
- 30 days
Scale
Govern
Register, own and govern agents, with approvals and audit evidence — before you enforce anything.
$2,465 / month
Annual commitment rate
Annual · was $2,900/mo
- Protected actions
- 250k / mo
- Environments
- 3
- Evidence retention
- 400 days
Card checkout for this is not switched on yet.
Runtime control
Enforce
Turn the same policies on. Allow, deny, redact, sandbox, approve or stop at the moment of action.
$7,565 / month
Annual commitment rate
Annual · was $8,900/mo
- Protected actions
- 1M / mo
- Environments
- 10
- Evidence retention
- 400 days
Card checkout for this is not switched on yet.
Scale
Enterprise
Residency, retention and deployment control for organizations that must prove it continuously.
$20,400 / month
Annual commitment rate
Annual · was $24,000/mo
- Protected actions
- 5M / mo
- Environments
- unlimited
- Evidence retention
- 2555 days
Card checkout for this is not switched on yet.
Discover is free and unmetered because finding an agent should never create a reason to hide it. Paid plans meter protected decisions, while enforcement continues even when a billing threshold is reached.
The platform
Five families of control over one normalized data model
Discovery, registry, identity, authority, enforcement, AI-native defenses and surface-specific controls — all reading and writing the same agent-authority-action graph.
Identity & Registry
Know every agent, who owns it, and what authority it actually holds.
5 controls →
Authority & Policy
Decide what an agent may do, under which conditions, with which credential.
4 controls →
Runtime Enforcement
Intercept the action itself: allow, deny, redact, sandbox, approve or stop.
4 controls →
AI-Native Security
Defend the parts of the stack that only exist because the actor is a model.
4 controls →
Agent Surfaces
Protect the specific places agents gain capability and take business action.
4 controls →
All 21 capabilities
Browse every control with its objective, build scope, definition of done and success metrics.
Open the index →
Families are how the controls group for reading. The product ships as a set of modules with two services underneath them, each one carrying the specification pages that define it.
Letting agents buy or pay? See the agentic commerce controls or run the free agent checkout risk check.
Trust model
Trust is earned by verification, never by location
Agent Trust Cloud sits in a privileged position, so its trust model is explicit by design: it minimizes implicit trust, makes decisions explainable, separates duties and provides verifiable evidence, rather than asking you to trust another AI model to keep your AI safe.
Agent trust levels
- 01
Unmanaged
Seen by nobody. No authority should flow here.
- 02
Observed
Detected in traffic or metadata, not yet claimed.
- 03
Registered
In the registry with an accountable owner.
- 04
Verified
Identity attested and bound to a policy context.
- 05
Privileged
Cleared for high-impact, policy-gated actions.
Principles we build to
- Identity before access: an agent must be attributable before privileged actions are allowed.
- Least privilege by default: grant only tools, actions, data scopes and time windows required for the task.
- Deterministic guardrails for irreversible/high-risk actions; AI can advise but not silently override hard policy.
- Continuous verification: reevaluate authority when context, owner, environment, model, tool or risk changes.
- Evidence by design: every control decision generates structured, exportable records.
How that is implemented
- Define trust levels for unmanaged, observed, registered, verified and privileged agents.
- Use explicit delegation chains when an agent acts for a person, service or another agent.
- Support policy simulation before enforcement to reduce operational risk.
- Design emergency access and break-glass procedures with mandatory logging and review.
TRUST MODEL
Never trust an agent because of where it runs; trust only verified identity, explicit authority, current context and policy.
Standards anchor
Built against published frameworks, not marketing claims
NIST SP 800-207
Zero Trust Architecture as the design reference for agent principals.
NIST AI RMF 1.0
AI risk management functions mapped to platform controls.
ISO/IEC 42001:2023
AI management system requirements.
OWASP Agentic Top 10
Excessive agency and agentic risk reduction.
EU AI Act
High-risk obligations apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
NIST 2026 concept paper
Software and AI-agent identity and authorization.
Learn
Plain-language references on AI regulation and governance
Why people are wary of agents
Every agent has an owner, a limit, a human approval and a record
58%
of executives said their company had an AI-related security issue or close call in the last 12 months.
19.7%
of organisations say all their agents are fully secured and governed before going live.
Gravitee, State of AI Agent Security 2026 (750 technology leaders)
7.2%
of organisations have a named individual formally accountable for AI agent behaviour.
Agent Trust Cloud closes those gaps: test every agent before release, give it a named owner and limits on what it may do and spend, hold consequential actions for a person, watch its behaviour and stop it with a kill switch, and keep evidence an auditor or insurer can sample. Check one of your agents free.
In brief
Agent Trust Cloud in four lines
- What it does
- A control plane for AI agents: it finds the agents you run, gives each an identity and owner, decides their actions against policy before they execute, and keeps the evidence.
- Who it is for
- Security, risk, compliance, platform and payments teams whose AI agents act on systems, data or money.
- Price
- Discover is free. The AI Governance Toolkit is $199 (Starter) or $599 (Professional), one-time. Platform plans are listed on the pricing page.
- Free option
- Start a free Discover workspace, or use the free agent checkout risk check and AI policy templates.
Where to start
Start by finding out what you already have.
A 90-day founding pilot finds the agents and machine identities running against one cloud account and one identity provider, gives each an owner, shows what each one can reach, and leaves you with audit evidence you keep. The fee is credited in full to your first year if you continue.
The control plane for trusted AI agents — identity, policy, runtime security, evidence and cost in one platform. See how it fits your stack.