CydraLabs

Zero-trust control plane for enterprise AI agents

Control every consequential action your AI agents take

Discover agents, map their effective authority, enforce policy before tools execute, and produce evidence for every decision across cloud, model and framework environments.

Request a demo See how the platform works

CydraLabs is the independent zero-trust control plane that discovers, authorises, monitors and proves every consequential action taken by enterprise AI agents.

Action decision

Illustrative
Agent
agent://acme/sales-agent
Delegated user
j.smith (Sales)
Tool
crm.export_contacts
Target
Production · Confidential
Agent Risk Score
59 · High
Inspection
Personal data detected

Decision

Approval required before execution

Evidence record appended to the hash-linked chain

Illustrative example of a CydraGateway decision record.

The problem

Enterprise agents are gaining authority faster than security can see it

Agents are built on many clouds, models and frameworks. Each one can reach tools and data through identities that security teams did not design for them.

Agents act with real authority

Agents now update CRM records, merge code, move money and send email. Each tool call is a consequential action taken on someone's behalf.

Authority is transitive and hidden

An agent's effective reach includes delegated user scopes, MCP servers and every tool those expose. It is rarely written down anywhere.

Controls were built for humans

Identity, approval and data-loss controls assume a person at a keyboard. Agent tool calls often bypass them entirely.

Evidence is missing when it matters

When an auditor or incident responder asks why an agent was allowed to act, logs rarely show the policy, context and approver behind the decision.

Platform

One control plane, shared by every product

CydraShield, CydraGateway and CydraGovern run on the same identity, policy, approval, telemetry and evidence services. Later products will build on them rather than duplicate them.

Explore the platform

  • Identity

    Unique agent identities, delegated-user binding and short-lived workload tokens.

  • Policy

    Versioned, deterministic rules evaluated before execution, with default deny.

  • Approval

    Named approvers and single-use, request-bound approval tokens.

  • Telemetry

    Activity metadata and content hashes by default, exported over OpenTelemetry.

  • Evidence

    Every decision appended to a signed, hash-linked evidence chain.

  • Graph

    Users, agents, models, MCP servers, tools, data and actions in one graph.

Core products

Launching with CydraShield, CydraGateway and CydraGovern

Every capability carries a status label: Available, Beta, Preview or Roadmap. Labels describe the current proof-of-concept build.

Security platform for AI agents

CydraShield

Discover every agent, understand its authority and score its risk.

CydraShield capabilities

  • Agent inventory and ownership

    Available

  • MCP server and tool discovery

    Beta

  • Model inventory (connector data)

    Available

  • Live model discovery from provider APIs

    Preview

  • Prompt and activity logging

    Available

  • Agent identity and permission mapping

    Available

  • Agent Risk Score (ARS)

    Available
All CydraShield capabilities

Secure agent, MCP and tool gateway

CydraGateway

Decide before tools execute — allow, deny or require approval.

CydraGateway capabilities

  • Identity-aware agent gateway

    Available

  • MCP, API and tool mediation

    Available

  • Pre-execution policy decisions

    Available

  • Human approval gates

    Available

  • Data and secrets inspection

    Available

  • Scoped credential handling

    Beta

  • Per-action evidence records

    Available
All CydraGateway capabilities

AI governance and compliance

CydraGovern

Turn runtime decisions into continuous compliance evidence.

CydraGovern capabilities

  • AI-system classification

    Beta

  • Control mapping across frameworks

    Beta

  • PCI DSS control coverage

    Roadmap

  • Evidence linked to runtime decisions

    Available

  • Exceptions and remediation tasks

    Available

  • Audit-readiness dashboards

    Available
All CydraGovern capabilities

CydraSOC, CydraDevSec, CydraCloud, CydraWorkforce, CydraKnowledge and CydraVoice are on the product roadmap.

Action control

Every consequential action is decided before the tool executes

CydraGateway puts a deterministic decision point between agents and enterprise tools. Ten steps, the same for every framework and model provider.

  1. Step 1: Receive the proposed action

    The agent submits the tool call to CydraGateway with an idempotency key before anything runs.

  2. Step 2: Authenticate the agent

    A signed, short-lived workload token is checked for audience, expiry, tenant and replay. Suspended agents are denied.

  3. Step 3: Resolve identity and target

    Owner, delegated user, tool, operation, target and environment are resolved. Unknown tools are denied by default.

  4. Step 4: Gather context

    Data classification, the latest Agent Risk Score and Agent Security Graph context are attached to the decision input.

  5. Step 5: Evaluate policy

    Versioned, deterministic policy is evaluated. A policy engine error fails closed.

  6. Step 6: Inspect content

    Parameters and payloads are checked for secrets, personal data and prompt-injection patterns.

  7. Step 7: Decide

    Allow, deny or require approval. Deny outranks approval, approval outranks allow, and no matching rule means deny.

  8. Step 8: Approve when required

    A named approver issues a single-use, time-bound token bound to the hash of the exact request.

  9. Step 9: Execute with scoped credentials

    The action runs with a per-execution credential limited to the target and a short lifetime.

  10. Step 10: Record evidence

    The decision and result are appended to the tenant's hash-linked evidence chain and signed.

Agent Security Graph

See the paths from people to agents to data

The graph connects users, agents, models, MCP servers, tools, data and actions, so you can answer questions that inventories alone cannot.

  • Which data can this agent reach, directly or transitively?
  • What is the blast radius if this MCP server is compromised?
  • What is the shortest path from an internet-exposed agent to confidential data?
  • How did this agent's access change over time?

Status: Available

Agent Security Graph example: a user delegates to a sales agent, which uses a model, connects to an unreviewed MCP server and can call a CRM tool. A highlighted path shows the agent reaching confidential customer data and an external email action through the unreviewed MCP server.delegatesusesconnectscan callexposesreadsreachesUserDelegating humanSales agentagent://acme/salesModelFoundation modelMCP serverUnreviewedCRM toolwriteCustomer dataConfidentialEmail sendExternal egressHigh-risk path (dashed)
Illustrative example, not customer data.

Solutions

Start from the outcome you need

Integrations

Connect the models, protocols and identities your agents already use

Connector adapters share one contract. In this build they run against mock data; live provider APIs are in Preview.

  • OpenAI

    Model provider

    Preview

    Discover assistants, models and tool definitions.

    Mock adapter Available; live API Preview

  • Anthropic

    Model provider

    Preview

    Discover models and agent configurations.

    Mock adapter Available; live API Preview

  • Model Context Protocol (MCP)

    Tool protocol

    Beta

    Enumerate MCP servers and tools via tools/list.

    Generic MCP discovery Beta

  • Microsoft Entra ID

    Identity

    Preview

    Map service principals, app registrations and delegated permissions.

    Mock adapter Available; live API Preview

  • GitHub

    Developer platform

    Preview

    Find agents and tokens in repositories and apps.

    Mock adapter Available; live API Preview

Planned integrations

Roadmap
  • Okta
  • Google (Gemini) and open-weight models
  • Microsoft Sentinel
  • Splunk
  • Google Security Operations
  • CrowdStrike
  • Microsoft Defender
  • AWS Security Hub
  • ServiceNow
  • Jira
  • GitLab
  • Salesforce
  • SAP
  • Microsoft 365
  • AWS, Azure and Google Cloud

Product names identify interoperability targets and do not imply partnership or endorsement.

Deployment and trust

Designed so security teams can verify it

These are design properties of the current build. They are not certifications or third-party attestations.

  • Modular deployment

    A modular monolith with separate web, API and worker units, packaged as containers and runnable with Docker Compose. Components such as the policy engine are swappable behind stable interfaces.

  • Tenant isolation

    PostgreSQL row-level security on every tenant table, enforced for a non-owner application role. Tenant context comes from the authenticated session, never the request body.

  • Hash-linked evidence chain

    Each evidence record includes the hash of its predecessor and an Ed25519 signature. Verification recomputes the chain and reports any break.

  • Fail-closed enforcement

    If the policy engine or a required detector fails, high-risk actions are denied. Fail-open is an explicit setting limited to read-class actions.

  • OIDC sign-in

    Portal sign-in uses OpenID Connect through a backend-for-frontend session with CSRF protection. Enterprise identity-provider templates are in Preview.

  • Least-privilege roles

    Nine roles from one permission matrix, enforced by the API. Policy authors cannot publish their own policies.

Design-partner programme

Design-partner programme

We are recruiting a small number of organisations to shape CydraShield, CydraGateway and CydraGovern with us. This section describes the programme; it is not customer evidence, and we do not display customer logos, quotes or results.

Apply to become a design partner

What design partners receive

  • Early access to the proof-of-concept platform
  • Direct input into roadmap priorities
  • Joint threat modelling of your agent estate
  • Working sessions with the engineering team

What we ask

  • Real agent use cases, in a non-production environment
  • Regular feedback sessions
  • A named security or platform sponsor
  • Candid assessment of what does not work

Resources and Academy

Learn, assess and go deeper

Put a control point in front of your agents' actions

See CydraShield, CydraGateway and CydraGovern working together on a realistic agent estate, or start with an indicative risk self-assessment.

Request a demoAssess your agent risk