Skip to content

AI security and governance

Securing and governing AI systems

AI applications bring new ways to be attacked and new rules to follow. I work on both: testing an AI system the way an attacker would, then showing it meets the frameworks a regulator or auditor will ask about.

Already delivered

What exists today

Delivered

Compliance Gate

A tool I built that tells a business what the EU AI Act requires of it, what to do to comply, and the one-time and ongoing cost.

Open the checker

In progress

Attack it, secure it, then govern it

One AI system, taken through six phases. Security comes first: I break it and fix it before assessing it against any framework.

Enterprise AI assistant: where I attack it

Built on made-up company data, so every result can be published.

  1. User prompt

    Prompt injection

  2. Login and roles

    Broken access control

  3. Document search (RAG)

    Poisoned or leaked documents

  4. Language model

    Sensitive data disclosure

  5. Agent and tools

    Excessive agency

  6. Logs and monitoring

    Undetected misuse

  1. Understand

    How language models, document search and agents work, and where the risk comes from.

    In progress

  2. Govern the data

    An AI risk register, a data inventory and a privacy assessment under GDPR and India's DPDP Act.

    Planned

  3. Attack

    Threat model, OWASP LLM Top 10 testing and a red-team report mapped to MITRE ATLAS.

    Planned

  4. Defend

    Guardrails, permission-aware document search, tool authorisation and an AI gateway, then a retest.

    Planned

  5. Govern

    Risk assessment with NIST AI RMF, gap assessment against ISO/IEC 42001 and EU AI Act classification.

    Planned

  6. Run the programme

    Governance charter, intake workflow, vendor questionnaire, incident playbook and an internal audit.

    Planned

Frameworks

The standards this work is measured against

OWASP Top 10 for LLM Applications
Prompt injection, excessive agency, data and model poisoning and the other main LLM risks.
MITRE ATLAS
Tactics and techniques for attacks on AI systems, used to label red-team findings.
NIST AI RMF
Govern, Map, Measure and Manage: the structure for an AI risk assessment.
ISO/IEC 42001
The certifiable management system standard for AI, with its Annex A controls.
EU AI Act
Risk tiers, provider and deployer duties, and the requirements for high-risk systems.
GDPR and India's DPDP Act
Lawful basis, people's data rights and when an impact assessment is needed.

Deliverables

What gets published

Each item is a document you will be able to open. The status shows where it stands today.

  • Architecture and data-flow diagramIn progress
  • AI risk registerIn progress
  • AI data inventory and privacy assessmentPlanned
  • AI threat modelPlanned
  • OWASP LLM Top 10 assessmentPlanned
  • AI red-team report, mapped to MITRE ATLASPlanned
  • Before and after control retestPlanned
  • AI supply-chain assessmentPlanned
  • NIST AI RMF risk assessment and model cardPlanned
  • ISO/IEC 42001 gap assessmentPlanned
  • EU AI Act applicability matrixPlanned
  • AI governance charter and acceptable use policyPlanned
  • AI incident response playbookPlanned
  • Internal audit report and framework crosswalkPlanned

Working toward the IAPP AI Governance Professional (AIGP) certification.