Skip to main content

Trust Center

Trust is an operating model, not a badge.

CLIP is designed to combine realistic cyber activity with enforceable authorization, independent safety controls, evidence integrity and accountable human decisions.

CLIP · Control trace

Illustrative view
  1. 01 · ObligationOrganizational or regulatory requirement
  2. 02 · Control objectiveWhat must be true
  3. 03 · ImplementationArchitecture or process control
  4. 04 · VerificationTest, review or assessment
  5. 05 · EvidenceRetained, dated, attributable

One catalog, not separate checklists per framework.

Applicability is deployment-dependent

Exact applicability, certification and accreditation depend on deployment scope, customer obligations and formal assessment. Nothing on this page asserts that CLIP holds a certification.

Trust principles

Four properties the product is built around.

  • Secure by design

    Identity, least privilege, segmentation, supply-chain assurance and auditability are default behaviours rather than deployment options.

  • Safe by design

    Offensive actions are authorized, isolated, bounded, monitored and terminable — enforced outside the workload performing them.

  • Evidence first

    Readiness claims trace to observed performance. Original evidence is immutable and independently verifiable.

  • Privacy aware

    Personal data is processed for declared purposes with defined basis, retention and access, and minimised wherever identity is unnecessary.

Safety invariants

Properties that hold regardless of scenario logic.

These are not procedural commitments. Each is enforced by a control that the exercise itself cannot reach.

  • An exercise can act only on resources bound to its authorization context.

    How it is enforced

    Target allowlists are evaluated outside the exercise workload. An action against an undeclared asset is denied at the network and policy layers, not by the tool requesting it.

  • Learner-controlled systems cannot change platform safety policy.

    How it is enforced

    Safety policy lives in the management zone, which range workloads have no route to. Policy decisions are made by a service the exercise cannot reach or modify.

  • Offensive agents cannot grant themselves tools, targets, credentials or routes.

    How it is enforced

    Tool access uses an allowlisted registry with least-privilege, time-bounded capability tokens issued externally. The model has no path to widen its own scope.

  • Emergency stop remains available when learner and exercise workloads fail.

    How it is enforced

    Stop operates through an independent control path that does not depend on the health of the target application or the exercise network.

  • Exercise completion cannot erase or alter original evidence.

    How it is enforced

    Originals are write-protected or stored in immutable object form. Analysis uses controlled working copies linked back to the original.

  • Scoring cannot silently change after adjudication.

    How it is enforced

    Adjustments preserve the original automated result, the assessor identity and the recorded rationale as separate, retained facts.

  • No AI recommendation becomes a credential or operational assignment without a human decision.

    How it is enforced

    AI outputs are labelled, explainable to the appropriate reviewer, and subject to human override. Credential issuance requires a separate authorized actor.

Zero Trust and segmentation

Network location alone never establishes trust.

  • All access is explicitly authenticated and authorized at the service boundary
  • Administrative and exercise-control interfaces require multi-factor authentication
  • Service-to-service communication across trust boundaries uses mutually authenticated encryption
  • Management, learning, exercise, malware, telemetry, evidence and exchange zones are separately controlled
  • High-risk exercise networks use default-deny egress with declared destination allowlists
  • Security controls fail closed when an authorization dependency is unavailable

Evidence integrity

Chain of custody, not just storage.

An evidence package must be verifiable by someone who does not have access to the platform that produced it. That constraint shapes how evidence is acquired, sealed and exported.

  1. 01Acquisition identifies collector, source, exercise, time and method
  2. 02Each artifact carries a SHA-256 or stronger approved hash
  3. 03The original is write-protected or stored as an immutable object
  4. 04Analysis uses controlled working copies linked to the original
  5. 05Every custody, export, transformation and access event is recorded
  6. 06Clock status and known time uncertainty accompany time-sensitive evidence
  7. 07Deletion follows approved retention, legal hold and destruction procedures

Agentic AI governance

The boundary is external to the model.

AI increases realism and personalisation. It never acquires authority, and it never becomes the decision-maker.

  1. 01Exercise authorizationValid context required to start.
  2. 02Policy decisionEvaluated outside the model.
  3. 03AgentRegistered, versioned, risk-classified.
  4. 04Allowlisted toolLeast-privilege, time-bounded token.
  5. 05Declared targetBound to the exercise only.

In sequence: An agent acts only within a valid exercise authorization, its request passes an external policy decision, it may use only allowlisted tools against declared targets, high-impact actions require human approval, and every step is recorded.

Human approval is not a footnote

High-impact actions require configurable human approval or dual control. Controllers can suspend one agent, all agents in an exercise, or the entire AI execution service.

Privacy and learner monitoring

Transparency about what is observed.

Training platforms observe people. That is only acceptable when learners know what is recorded, why, and for how long.

  • Learners receive understandable notices for monitoring, recording and automated analysis
  • Personal data is processed for declared purposes with a defined basis and retention
  • Analytics minimise personal data and use pseudonymisation or aggregation where identity is unnecessary
  • A privacy impact assessment is completed before behavioural analytics, proctoring or AI assessment is enabled
  • Correction workflows preserve audit history rather than silently rewriting evidence
  • Research datasets are de-identified and approved before release

Standards and obligations

Designed to support, not certified against.

Standards below are engineering and governance inputs. Applicability must be confirmed for each customer and deployment; they are not automatically legally mandatory.

  • UU No. 27/2022 (PDP)

    Conditional

    Indonesian obligation

    Personal data protection. Requires legal applicability review for the deployment jurisdiction.

  • Peraturan BSSN

    Conditional

    Indonesian obligation

    Sector and national cybersecurity requirements where they apply to the operator.

  • ISO/IEC 27001

    Guidance

    Management system

    Organizational information security management context. Certification is held by an organization, not by software.

  • ISO/IEC 27701

    Guidance

    Management system

    Privacy information management context for the operating organization.

  • ISO/IEC 42001

    Guidance

    AI governance

    AI management system context for agentic capabilities.

  • NIST SP 800-218 (SSDF)

    Baseline

    Engineering baseline

    Secure development lifecycle practices applied to the product.

  • OWASP ASVS

    Baseline

    Engineering baseline

    Application verification baseline for hardened surfaces.

  • NIST SP 800-207

    Baseline

    Engineering baseline

    Zero Trust architecture principles across trust zones.

  • NIST NICE / SKKNI

    Conditional

    Workforce framework

    Competency mapping targets. CLIP maps to them; it does not own them.

What these statements mean

CLIP is designed to support and is aligned with the practices above. It supports competency mapping to organizational, NICE and relevant SKKNI frameworks. Formal certification can be delivered with authorized certification partners. No statement here asserts that CLIP is approved or certified by any authority.

In summary

Four boundaries that do not move.

  • Declared targets and Rules of Engagement

    No exercise, offensive tool or AI agent starts without approved Rules of Engagement and an explicit target scope.

  • Default-deny isolation and independent emergency stop

    Range egress is denied by default, and emergency stop operates independently of the learner-controlled environment.

  • Immutable evidence and chain of custody

    Original evidence is write-protected and integrity-verifiable; analysis happens on controlled working copies.

  • Human authority over AI, scoring and release

    AI recommendations and platform scores support authorized human decisions. They never independently determine an outcome.

Responsible disclosure

Reporting a security issue.

Replace before publishing

A responsible-disclosure contact address, PGP key, acknowledgement window and safe-harbour statement must be confirmed by the operating organization before this site is published. Do not publish an unverified security contact.

Request an assurance briefing.

We will walk through the safety invariants, evidence model and control mapping against your obligations.

Security, Safety and Trust | CLIP