Skip to main content

Capability

Adaptive cyber training. Bounded by policy. Accountable to humans.

Use registered AI adversaries, defenders, pentesting assistants and learning coaches only inside declared training contexts with allowlisted tools, target binding, complete action traces and independent stop controls.

CLIP · Agent policy and approval trace

Illustrative view
  • Agentadversary-emulation · v2.1 · risk class B
  • Declared targetEX-2026-014 range segment only
  • Policy decisionAllowed with human approval
  • Tool requestAllowlisted · capability token 15 min
  • ControlsSuspend agent · Stop exercise agents

Illustrative agent trace. Human approval is required before the high-impact action proceeds.

CLIP does not provide unbounded autonomous offensive capability or authorize activity against external or production targets.

Mission problems

What this capability addresses.

  • Scenario realism does not adapt to the learner's actual level
  • Personalised coaching does not scale across a large cohort
  • AI tooling is adopted without enforceable boundaries or an audit trail
  • Sovereign deployments cannot depend on an external inference service
  • AI-generated assessment is used without human review or explainability

Capability modules

What is included.

Modules are composable. A pathway combines the ones a role actually needs.

  • Adaptive adversary for Blue Team training
  • Defensive assistant for Red Team challenge and analysis
  • Pentesting assistant restricted to approved target environments
  • Learner coach for hints and personalised progression
  • Scenario generator assistant with human review
  • After-action reconstruction assistant

End-to-end workflow

How the work flows.

  1. 01Exercise authorization
  2. 02Policy decision
  3. 03Agent action
  4. 04Allowlisted tool
  5. 05Declared target
  6. 06Recorded trace

In sequence: An agent may act only inside a valid exercise authorization context. Every tool request passes an external policy decision, may act only through allowlisted tools against declared targets, and produces a complete recorded trace including any human approval.

Representative scenario

One scenario, end to end.

A bounded AI adversary escalates activity against a declared training network while a coach agent offers the Blue Team hints. Every high-impact tool request pauses for human approval, and the controller can suspend one agent or all agents in the exercise.

Evidence produced

  • Prompt, context reference, model and version
  • Tool request with the policy decision that allowed or denied it
  • Declared target binding for every action
  • Human approval record for high-impact actions
  • Evaluation results for boundary adherence and harmful behaviour

Evidence and metrics

What you can measure.

These are the operational measures the platform produces. Baselines and targets are set with each organization during a pilot rather than claimed in advance.

  • Boundary adherence in adversarial evaluation
  • Task success against the training objective
  • Assessor agreement with AI-generated judgements
  • Emergency-stop and agent-suspension time

Safety and trust

The boundary is external to the model.

  • Agents act only on declared asset identifiers bound to the exercise

  • Tool access uses least-privilege, time-bounded capability tokens

  • High-impact actions require configurable human approval or dual control

  • Resource, time, action-rate and network limits are enforced outside the model

  • Agents cannot self-replicate, persist, modify safety controls or widen their own privileges

  • Local and offline inference is supported for sovereign deployments

See the full trust model

See this capability against your mission.

We will tailor the demonstration around your priority use case and operating constraints.

Agentic AI Training | CLIP