UKVPM Technology & R&D Programme
COREHAVEN™
AI, Resilience & Applied R&D
Corehaven is a UKVPM Ltd technology and R&D programme developing resilient digital systems for risk awareness, preparedness and the safe integration of artificial intelligence.
Within the programme, Corehaven Guardian is an evidence-driven assurance layer for evaluating and governing AI-assisted and agentic operations in security- and resilience-sensitive environments. Its core controls are a working foundation, under continued development and validation.
A working platform for applied research
Corehaven is not only a concept. Its development platform combines risk information, preparedness tools and AI-assisted guidance in one environment, giving the programme somewhere real to research, build and demonstrate AI and resilience capabilities.
It is an active R&D platform and demonstrator environment. It has not been deployed by government or public-sector organisations.
Preparedness assessment
Structured readiness assessments that help users understand how prepared they are and what to do next.
Climate adaptation and risk scoring
Scoring that relates climate-related risk to adaptation and preparedness measures.
Location-aware risk information
Risk information presented in the context of a user's location.
AI-assisted guidance
AI-assisted preparedness guidance, designed to direct safety-critical situations towards emergency support and human help rather than relying on the model alone.
The problem
AI agents that can act need more than chatbot safety
Increasingly capable AI agents can access information, use tools, call APIs, run workflows, make recommendations and initiate actions with real consequences.
Controlling what a model says is no longer enough. Organisations need a reliable way to understand and govern what an agent is about to do, before it does it, and to account for it afterwards.
Questions an organisation needs to answer
- 1Which agent is operating, and on whose behalf?
- 2What is it attempting to do?
- 3Which tools and data can it reach?
- 4What evidence supports the proposed action?
- 5Is the action permitted by policy?
- 6How risky is it in this specific context?
- 7Does a person need to decide?
- 8What happened afterwards, and can it be shown?
COREHAVEN Guardian
Evidence-based AI agent assurance
Guardian is a controlled assurance layer between AI agents and the tools, data and systems they act on. Rather than trusting an agent's own judgement, it assesses each proposed operation against evidence, context and policy, and keeps an accountable record of the outcome.
Guardian's decisions are driven primarily by deterministic policy and risk controls. AI can contribute to an assessment where appropriate, but it does not replace those controls.
- 01
AI or agent request
An AI-assisted system or agent proposes an operation, such as calling a tool, retrieving data or starting a workflow.
- 02
Context and evidence
The request is gathered with its operational context and the evidence it relies on.
- 03
Guardian assessment
Evidence is checked, input is screened for manipulation and contextual risk is assessed.
- 04
Risk and policy controls
Guardian policies, permissions and data boundaries are applied deterministically.
- 05
Decision
The operation is allowed, blocked, quarantined or held for human review. Blocked and quarantined operations raise an incident.
- Allow
- Block
- Quarantine
- Human review
- 06
Controlled action
Only an authorised operation proceeds to the tool or system concerned.
- 07
Auditable record
The assessment, evidence, decision and outcome are kept as a tamper-resistant record for review and accountability.
Built for both security and resilience
Safe agent adoption depends on two things at once: resisting deliberate misuse, and behaving predictably when things go wrong without any attacker involved. Guardian is designed to address both.
Cybersecurity
- Malicious or manipulated inputs
- Prompt injection
- Excessive permissions
- Unauthorised tool use
- Agent identity
- Policy enforcement
- Data boundaries
- Evidence integrity and auditability
Operational resilience
- Unavailable or degraded services
- Incomplete or conflicting evidence
- Model or provider failure
- Safe fallback behaviour
- Escalation when confidence is low
- Operational continuity
- Failure containment
- Recovery
Human oversight by design
Guardian is not intended to replace human decision-making. Its role is to make sure the right decisions reach the right people, with the evidence they need, and that accountability is preserved. The core approval and escalation mechanism is a working foundation; broader organisational approval workflows are in active development.
- Risk-based escalation
- Approval for consequential operations
- Human-in-the-loop pathways
- Human-on-the-loop monitoring
- Decisions recorded with their evidence
- Auditable, accountable operation
Independent of any single AI provider
Model ecosystems change quickly. Guardian and Corehaven's controlled AI access layer are being developed to remain independent of any single model provider, applying the same policies, evidence controls and audit record whichever AI service is in use.
Current capability and development status
Guardian's core controls are working foundations within the Corehaven development platform, supported by automated functional and security testing. They are not a finished commercial product or a certified service, and development and validation continue.
Each capability is labelled by its current status and described at a conceptual level only.
- Working foundation
- Implemented in the current development platform and exercised in testing. Not a claim of production certification.
- In active development
- Partly built and being extended, integrated or hardened.
- Research objective
- A question we are investigating. Outcomes are not yet established.
- Potential application
- A context where the work may be applied in future. Not a current deployment.
Working foundations
Agent and tool governance
A structured registry and control layer defining which AI-assisted systems and tools may operate, and in which authorised contexts.
Evidence validation
Operations are linked to identified evidence, and unsupported or invented evidence is rejected or controlled.
Evidence trust classification
Evidence is assigned a trust level that informs how far a decision may rely on it.
Prompt-injection detection
Guardian safety checks identify prompt injection and other manipulated input before an operation proceeds.
Contextual risk assessment
Each proposed operation is assessed against the agent, the tool, the evidence and the operational context.
Policy-based decisions
Deterministic, policy-driven decision logic across multiple Guardian policies. AI may inform an assessment but does not replace these controls.
Allow, block and quarantine
Every assessed operation receives an explicit outcome. Blocked and quarantined operations raise an incident for review.
Human approval and escalation
Consequential or uncertain operations can be held for a person to approve before they proceed.
Tamper-resistant decision records
Guardian assessments and decisions are kept as tamper-resistant, integrity-protected records designed to resist later rewriting.
Append-only evidence
Evidence is held in an append-only form, so the basis for each decision remains traceable.
Controlled AI access
AI services are reached through a centralised access layer with model and provider abstraction, and consequential operations require Guardian authorisation.
Controlled model access and routing
Model allowlisting and policy-governed, risk-aware routing, with fallback handling, provide a foundation for applying consistent organisational controls across different AI services.
AI operational logging and oversight
AI and Guardian activity is logged, with administrative oversight of operations and decisions.
In active development
Multi-agent governance
Extending governance from individual agents to agents that interact with one another.
Behavioural risk analysis
More sophisticated analysis of how an agent behaves over time, not only what it requests.
Tool-chain analysis
Assessing risk across sequences of tool calls rather than each call in isolation.
Advanced evidence provenance
Richer tracing of where evidence came from and how it has been handled.
Organisational approval workflows
Public-sector-style approval routes, delegation models and multi-level governance.
Cross-organisational policy
Policy structures that can be shared or layered across organisations.
Assurance for security leaders
Deeper automated assurance and reporting to support risk-based decisions by security leaders.
Resilience simulation
Exercising behaviour under degraded, failing or conflicting conditions.
Expanded adversarial testing
Broader and more systematic testing of controls against deliberate misuse.
Wider integrations
Connecting Guardian controls to a wider range of tools, services and environments.
Open research questions
Research objectiveSome of the most important questions in agent assurance remain unresolved across the field. Current R&D explores:
- Assurance for large-scale multi-agent systems
- Trust interoperability between organisations
- Standardised assurance for public-sector AI agents
- Advanced, explainable models of agent risk
- Systematic resilience evaluation across heterogeneous AI ecosystems
- Reusable and open assurance frameworks
Designed for high-trust environments
Potential applicationCorehaven's combination of resilience services and AI assurance is relevant wherever AI must operate within clear accountability. Potential application contexts include:
- Local government
- Public-sector AI services
- Emergency management and resilience planning
- Citizen-facing digital services
- Internal organisational AI agents
- Regulated or security-sensitive organisations
These are potential applications under exploration, not current deployments.
The R&D team
Corehaven is multidisciplinary. The programme draws on UKVPM's experience across digital platforms, artificial intelligence, media technology, cybersecurity, systems thinking and resilience.
Corehaven is one of the ventures of UKVPM founder Ali Jam.
Research leadership and academic experience
Corehaven's multidisciplinary R&D work is supported by Dr Zainab Saleh, who leads Research, Innovation and Strategic Partnerships at UKVPM Ltd and is a Senior Lecturer in Aerospace Engineering and Faculty Bidding Champion at Kingston University London.
Her research spans aerospace engineering and multidisciplinary innovation, bringing relevant expertise in complex systems, engineering research, resilience and technical validation. She currently co-leads a multidisciplinary project focused on supporting SMEs in achieving net-zero emissions.
Her UKVPM role is independent of her university appointment. Her Kingston University position is held independently and does not imply institutional participation or endorsement of Corehaven by Kingston University.
Collaborate on COREHAVEN
We welcome conversations with organisations, researchers and technology partners interested in AI assurance, resilience and evaluation of the programme. Further technical detail is shared with evaluators and partners under appropriate arrangements.
Corehaven is an active R&D programme under continued development and validation. Capabilities are described conceptually and labelled by status. This page is not a statement of certification, accreditation or production deployment.