Data security posture management, or DSPM, is the practice of continuously finding where sensitive data lives, who and what can reach it, and how exposed it is, then fixing the gaps. For AI compliance it answers a basic question regulators ask: which data went into this system, and was it allowed?
Is DSPM a defined standard?
No. DSPM is an industry term, and we found no entry for it in NIST's glossary on 18 October 2026. Treat it as a description of a set of activities, not a certification or a product category with fixed edges. That matters when you read a questionnaire or a vendor claim: ask which activities are meant.
The activities are familiar from older practice. Discover data stores. Classify what is in them. Map permissions. Watch for drift. What changed is scope. Data now sits in object storage, SaaS exports, vector databases, notebooks and prompt logs, often copied by people who never filed a ticket.
Why does AI make the data question sharper?
Models are built from data and then fed more data at run time. Each step creates a copy.
The EU AI Act, in Article 10, says training, validation and testing data sets for high-risk AI systems must be subject to data governance and management practices appropriate for the intended purpose. The NIST AI Risk Management Framework, released on 26 January 2023 and voluntary, frames the same concern as a risk to manage across the AI lifecycle. Neither document tells you to buy a tool. Both expect you to know your data.
Whether a system counts as high-risk is a legal question for your counsel. The data inventory work is useful either way.
What does a constructed example look like?
Imagine a support team that exports six months of chat transcripts to fine-tune an assistant. Nobody checks the export. Some transcripts contain card numbers and home addresses, because customers paste them in.
A posture review would find the export bucket, flag that it holds personal data, show that a contractor group can read it, and note that the fine-tuning job copied it a second time. Four findings, none of them exotic. Without the review, the first sign is a data subject request you cannot answer.
How to check or apply this
- List every place data enters an AI workflow: training sets, retrieval indexes, prompt and response logs, evaluation sets.
- Run discovery on each store and label data by sensitivity. Start with personal data, credentials and anything under contract.
- Map who and what can read each store, including service accounts and pipelines, not just people.
- Record the source and purpose of each training set. If you cannot say why you hold it, that is a finding.
- Set rules for what may not enter a model: raw secrets, unredacted identifiers, data collected for another purpose.
- Re-scan on a schedule. Posture is a rate of change, not a snapshot.
- Keep the evidence. An auditor wants dated outputs, not a policy that says you do this.
What about retrieval and prompt data?
Training is half of it. Many AI features retrieve documents at query time and log prompts and answers. Those indexes and logs are data stores like any other, and they often hold more sensitive material than the training set, because users type real questions into them.
Treat a retrieval index as a copy of its source. If a document is deleted or access to it is restricted at the source, the index needs to follow. Test this with one document: remove it, then ask the assistant about it. If the answer still quotes it, your deletion path has a hole.
Prompt logs deserve a retention rule. Decide how long you keep them, who can read them, and whether they may be reused for evaluation. Writing that down is cheap, and it is the kind of record an assessor asks for.
Where do teams go wrong?
They scan production databases and skip the copies. Most exposure lives in exports, backups and notebook output folders.
They treat labels as truth. A classifier that says a column is non-sensitive is a hint, not a ruling, so sample and verify the important stores.
They stop at discovery. Finding 40,000 exposed files is not progress unless someone owns closing them.
Quick check
Does DSPM make us compliant with the EU AI Act?
No. It helps you produce evidence of data governance, which is one part of the obligations for high-risk systems. Compliance depends on your role, your system's risk class and many other duties. Ask legal counsel which apply.
What is the first step if we have no data inventory?
Start with the stores that feed AI systems and with personal data. A small, accurate inventory of ten stores beats a large guessed one.
Sources
- EU AI Act, Article 10: Data and data governance
- NIST AI Risk Management Framework
- NIST CSRC glossary
Last reviewed: 18 October 2026.
