CydraLabsZero trust for AI agents

CydraRedTeam

Test how agents behave under attack before they ship

Agents ship without evidence of how they behave under prompt injection, tool or MCP poisoning, memory poisoning, excessive agency or privilege escalation.

RoadmapAssess your agent risk

The problem

What CydraRedTeam resolves

  • Prompt injection

    Instructions hidden in content an agent reads can redirect what it does.

  • Tool and MCP poisoning

    A tool description or MCP server can steer an agent towards harmful calls.

  • Memory poisoning

    Content written to an agent's memory can change its later decisions.

  • Excessive agency and privilege escalation

    Agents can reach more authority than their task needs, directly or through delegation.

Where it acts

Where Cydra controls an agent action

CydraRedTeam runs campaigns against agents connected through CydraGateway, so every test action is decided, recorded and contained.

Action-control workflow: receive, authenticate, resolve, context, policy, inspect, decide, approve, execute with scoped credentials, evidence.1Receive2Authenticate3Resolve4Context5Policy6Inspect7Decide8Approve9Execute10EvidencePolicy errors fail closed · Deny outranks approval outranks allow · Every outcome produces an evidence record
The ten-step action-control workflow shared by every product.
Step-by-step explanation

Capabilities

CydraRedTeam capability tour

  1. Discover

    • Scenario-based campaigns

      Roadmap

      Campaigns built from attack scenarios against agents, MCP servers and tool chains.

  2. Decide

    • Recommended policies

      Roadmap

      Findings become recommended CydraGateway policies to review and publish.

  3. Enforce

    Provided by the core platform.

  4. Investigate

    • Findings in the Security Graph

      Roadmap

      Each finding is attached to the agent, tool or MCP server it affects and feeds the Agent Risk Score.

  5. Prove

    • Evidence

      Roadmap

      Campaign actions and results recorded in the evidence chain.

Approach

Scenario-based campaigns, not a one-off scan.

  • Choose the agents and MCP servers in scope from the CydraShield inventory
  • Run scenarios for prompt injection, tool and MCP poisoning, memory poisoning, excessive agency and privilege escalation
  • Review findings on the Security Graph and the policies they recommend

Available today

Discovery, effective-authority analysis, prompt-injection monitoring and fail-closed gateway policy already reduce these risks in production.

CydraGateway

Coverage

Integrations and deployment

Every item carries its current status.

Integrations

  • Agents connected through CydraGateway

    Campaign targets

    Roadmap

  • MCP servers discovered by CydraShield

    Campaign targets

    Roadmap

Deployment

  • CydraLabs-managed SaaS

    Hosted and operated by CydraLabs.

    Pilot programme

  • Customer cloud or private VPC

    Deployed into your own cloud account with the CydraLabs reference deployment.

    Pilot programme

  • On-premises or sovereign deployment

    Planned for environments that cannot use cloud services.

    Roadmap
Full capability list with status

  • Scenario-based red-team campaigns

    Campaigns that test agents, MCP servers and tool chains for prompt injection, tool and MCP poisoning, memory poisoning, excessive agency and privilege escalation.

    Roadmap

  • Findings mapped to the Security Graph

    Each finding is attached to the agent, tool or MCP server it affects and feeds the Agent Risk Score.

    Roadmap

  • Recommended gateway policies

    Findings become recommended CydraGateway policies to review and publish.

    Roadmap
Available
Implemented, tested and demonstrable in the current proof-of-concept build.
Beta
Implemented and demonstrable, with documented limitations.
Preview
Interface or mock implementation behind a real contract; live integration not yet enabled.
Roadmap
Not built in this phase. Shared platform foundations are in place.

Trust

Security, privacy and trust

  • Campaigns run against agents you register, through the same policy and approval controls as production actions.
  • Tenant isolation with PostgreSQL row-level security for the application role.
  • Payloads stored as hashes plus redacted previews by default; no secrets in logs.
  • TLS in transit; encrypted storage in the AWS reference deployment.
  • No third-party certifications or attestations yet. See the Trust Centre for current assurance status.
Visit the Trust Centre

Proof

What we can show today

  • Roadmap status

    CydraRedTeam is on the roadmap. It is not built in the current phase, and we do not publish coverage or benchmark figures for it.

  • Design-partner outcomes

    Pilot outcomes and case studies will be published when design partners agree. We do not publish customer names, logos or results before then.

Find out where your agents are exposed today

Start with an indicative agent risk assessment, or join the pilot to shape CydraRedTeam.

Assess your agent riskJoin the pilot