For CISOs, CIOs and AI product owners
Security and
AI risk management
We help companies implement AI with clear rules, limits and responsibilities. We assess data, model and agent risks, design safeguards and define how AI systems are approved for production.
Scope of workProject scope
- RisksThreat scenarios and business consequence assessment
- ControlSafeguards and effectiveness testing
- GovernanceResponsibility, approval and change control
Business challenges
Risk control throughout the AI lifecycle
We work with new projects and existing solutions, linking technical controls to specific consequences for the company.
Discuss an assessmentEmployees already use external AI services
You need an inventory of usage, data handling rules and an approved tool list.
AI accesses corporate knowledge and actions
You need to verify retrieval permissions, API access, permitted operations and human confirmation of critical actions.
The pilot is ready to enter production
You need approval criteria, safeguard testing, risk owners and monitoring and response procedures.
Architectural approach
Protection at data, model and action levels
A request filter alone cannot address every security issue. We establish several control layers and test them against scenarios for the specific system.
The layers are interconnected.
We align decisions within the overall architecture.
Data and access
We check which data enters the AI system and who can obtain it through answers, retrieval or logs.
- Data and documents
- Permissions and filtering
- Permitted context
Key decisions
- Classification, approved processing environments and retention periods
- Source and RAG retrieval permissions, user and department isolation
- Secrets, personal data and sensitive information in logs
An access matrix and data handling requirements for each scenario.
Models and answers
We evaluate system behaviour under malicious instructions, unreliable context and incorrect answers.
- Request and context
- Model / checks
- Validated output
Key decisions
- Direct and indirect prompt injection and attempts to reveal sensitive data
- Processing outputs before passing them to interfaces, code and business systems
- Model provenance, dependencies, updates and provider terms
A threat model, test scenarios and safeguard requirements.
Agents and tools
We constrain agent actions through application and infrastructure controls, beyond textual model instructions.
- Agent intent
- Permission check
- Action / confirmation
Key decisions
- Least privilege, permitted tools and argument validation
- Human confirmation of operations with significant consequences
- Cost and request limits, isolated execution and emergency stopping
An agent permissions matrix and rules for executing and confirming operations.
Governance and monitoring
We assign risk ownership and define post-launch control procedures.
- Register and owner
- Approval / monitoring
- Review and change
Key decisions
- An AI system register, risk classification and risk acceptance owners
- Control over changes to models, prompts, data and tools
- Monitoring, logging, response and reassessment
An AI Governance process and AI system approval and support procedures.
Let's discuss your architecture
We will review the current situation, constraints and target use cases.
Scope of work
From technical decisions
to a working system
We agree the assessment scope with IT, information security, process owners and legal teams. Review depth depends on data, system authority and the consequences of possible errors.

Inventory and risk assessment
We identify where the company uses AI and which risks need priority controls.
- A register of AI systems and providers, use cases, data categories and owners.
- Threat modelling based on architecture, integrations and trust boundaries.
- Assessment of likelihood, consequences and residual risk, with a treatment plan and owners.
Stage outcome
A risk register, threat model and prioritised action plan.

Protecting data and corporate knowledge
We review the entire data path from source to user answer.
- Data classification and minimisation, masking, storage and deletion.
- RAG access controls across sources, indexes, retrieval and responses, including permission revocation.
- Assessment of external services, network calls, telemetry and logs.
Stage outcome
Data handling requirements, an access matrix and a gap remediation plan.

LLM and AI agent security
We design constraints appropriate to the system's authority.
- Protection against prompt injection and unsafe output handling, with untrusted content checks.
- Tool and API access controls, parameter validation and critical action confirmation.
- Dependency and model controls, execution isolation, quotas and secret management.
Stage outcome
Safeguard architecture and implementation and configuration requirements.

Testing protection effectiveness
We test controls against an agreed set of normal and adverse scenarios.
- A test programme covering permission bypass, data leakage, prompt injection and unwanted agent actions.
- Testing within an authorised environment using agreed test data and restrictions.
- Result analysis, remediation priorities and retesting after changes.
Stage outcome
A report with reproducible checks, consequence assessment and recommendations.

Enterprise AI Governance
We integrate AI governance into existing company processes.
- AI usage policy, permitted scenarios and tools, and data handling rules.
- Business, IT, information security and legal roles; approval and residual risk acceptance procedures.
- Production approval criteria, provider management and change control.
Stage outcome
A policy package, responsibility matrix and AI lifecycle procedures.

Monitoring and response
We define how deviations are detected and managed after launch.
- Events to monitor, alerts, action audits and integration with existing security processes.
- Response scenarios: restricting access, stopping tools, rollback and recovery.
- Periodic risk reassessment, owner training and monitoring of action completion.
Stage outcome
Monitoring procedures, response scenarios and a regular review plan.
Agree on the project scope
We will define the work required, participants and success criteria.
Consulting deliverables
Documents for decisions
and implementation
We agree on the documentation package at the start. The materials must support agreement with business, IT and security stakeholders and the engineering team's subsequent work.

AI and risk register
Systems, owners, scenarios, consequences and priorities.
Threat model
Trust boundaries, attack surfaces and applicable threat scenarios.
Protection architecture
Controls, access, agent restrictions and data protection.
Test report
Conditions, observations, remediation priorities and retesting.
Policies and approval
Usage rules, roles and system launch criteria.
Support plan
Monitoring, incidents, changes and risk reassessment.
Project organisation
We work together
with your team
At each stage, we agree on decisions and next actions. Timelines depend on scope, availability of source data and the depth of technical checks.
- 01
Define the scope
We select systems, data and scenarios, agreeing access and testing rules with owners and security teams.
- 02
Assess risks
We study architecture, processes and authority, then build a threat model and priorities based on business consequences.
- 03
Design and test controls
We prepare requirements and carry out agreed tests, documenting limitations and residual risks.
- 04
Embed controls in operations
We agree owners, approval criteria, monitoring and a remediation plan, then hand procedures to the responsible teams.
What is needed
to get started
- List of AI systems, models and providers
- Data, integration and agent authority diagrams
- Internal policies and security requirements
- System owners and permitted testing scope
Consulting and technical assessment are not certification or a legal opinion. Applicable mandatory requirements are agreed with the client's legal and information security teams. Residual risk acceptance remains the risk owner's decision.
Before the project starts
Questions and answers
Does local hosting guarantee AI security?
No. On-premises hosting helps control location and infrastructure access, but does not eliminate permission errors, injection risks, unsafe agent actions or integration vulnerabilities. We assess the whole system.
Which approaches do you use?
Our assessment uses applicable NIST AI RMF and OWASP practices for LLMs and generative AI, alongside the client's internal requirements. Checks are adapted to architecture and scenarios rather than treated as a universal security certificate.
Can we start with an existing RAG assistant?
Yes. We check document and retrieval permissions, untrusted content handling, answers and logs. We agree test accounts and data, then document gaps and a remediation plan.
How do you control an agent that modifies CRM or ERP data?
We define permitted operations, minimum permissions and application-side argument validation. Critical actions include human confirmation, auditing and a stop mechanism. Specific controls depend on the operation's consequences.
Do you test in production?
The environment and permitted actions are agreed before work begins. We prioritise an isolated environment with representative test data. Production testing is possible only within separately approved boundaries and with the responsible teams involved.
Next step
Define the risks and requirements for your AI project
In a working session, we will clarify scenarios, data and system authority. We will agree the assessment scope, participants and deliverables needed for a launch decision.
