Ghani Governance · Microsoft Purview
AI data security threats
AI data security covers the ways generative AI can expose or corrupt organisational data: sensitive data pasted into prompts or uploaded files, prompt injection and jailbreaks that extract information, unapproved “shadow AI” tools, and poisoned training or grounding data. Ghani Governance maps each threat to Microsoft Purview controls such as DLP, sensitivity labels, DSPM for AI, Insider Risk and audit, as part of a practical AI governance programme.
The threats we see in AI security reviews
Data loss via prompt and upload
Users paste customer or commercial data into AI tools, or upload whole files. Purview endpoint and browser DLP can warn or block for supported AI sites.
Prompt injection
Instructions hidden in a document, email or web page cause an AI system to ignore its rules and disclose data (OWASP LLM01). Labels and least-privilege access limit what can leak.
Jailbreaks
A jailbreak (including the widely shared “ChatGPT jailbreak” prompts) tries to bypass a model’s safety rules. Communication Compliance’s Copilot interactions policy template includes the Prompt Shields classifier, which flags jailbreak and prompt-injection attempts, where licensed.
Shadow AI and shadow IT
Shadow IT is any technology used without IT approval; shadow AI is the same problem with AI apps. DSPM for AI and Defender for Cloud Apps show which AI sites people use.
Data poisoning and data drift
Data poisoning deliberately corrupts training or grounding data; data drift is gradual change that degrades outputs. Both are data governance problems before they are model problems.
AI hallucination
An AI hallucination is a confident but false output. It is a quality and liability risk, especially when outputs are reused without review.
What is shadow IT, and why does shadow AI matter?
What is shadow IT? Apps, services and devices used for work without approval. Shadow AI matters because the data leaves with the prompt. The answer is rarely a blanket ban: approve a route (for example Microsoft 365 Copilot Chat under your commercial terms), then apply DLP to the rest.
Is ChatGPT safe for work data?
Is ChatGPT safe? It depends on the version, the contract and the data. Consumer accounts, ChatGPT Enterprise and Azure OpenAI have different data-handling terms. Without DLP, any of them can receive data it should not. See Copilot data security.
AI governance: responsible AI in practice
“AI transformation is a problem of governance” is a phrase that circulates for good reason: the hard part is ownership, policy and evidence. An AI policy states what tools are approved and what data may be used; AI risk management and AI compliance then check it is followed; responsible AI principles set the bar for fairness and transparency. AI governance tools in Microsoft Purview include DSPM for AI, Compliance Manager AI templates, Communication Compliance and audit. AI red teaming – structured adversarial testing of an AI system – is a separate specialist service; we can scope the data controls it tests. AI cybersecurity more broadly includes identity and network controls we do not deliver. See AI security frameworks.
Does the UK have an AI policy, and what does AI compliance involve?
The UK has a national AI policy (including the AI Opportunities Action Plan published in January 2025) but no single AI statute. AI compliance for a UK organisation therefore means existing law, chiefly UK GDPR, equality and sector rules, plus the EU AI Act where it applies. An organisational AI policy should name approved tools, the data that may and may not be used, and who approves new AI use cases. Purview DLP, DSPM for AI and audit then show whether the policy is followed.
AI threat questions
Can Purview stop prompt injection?
Not on its own. Prompt injection is a model and application problem. Purview reduces the impact by limiting what data the AI can reach and logging what it did.
Do you do AI red teaming?
No. We do not run adversarial model testing. We secure the data around AI systems and can work alongside a red team.
Book a discovery workshop
A scoping conversation about what you need to protect, what is already configured, and what must not break.