Topic starter
17/08/2026 4:37 pm
TL;DR: Fixed-pattern DLP misses sensitive data hidden in ordinary AI conversation, while WitnessAI’s NER-D uses double-pass named entity recognition to classify meaning at runtime, not just shape, and its benchmarked method was accepted at ACL 2026. The security issue is not detection volume alone, but whether identity-like data, secrets, and regulated content can be governed when context determines sensitivity.
Editorial analysis by NHI Mgmt Group, based on content published by WitnessAI: “Why we built NER-D”.
Key questions
Q: Why do DLP and allow/block controls fail for conversational AI?
A: Because they look for known patterns and static destinations, while conversational risk depends on meaning, purpose, and session context.
Q: How should security teams handle sensitive data in enterprise AI chats?
A: Security teams should treat enterprise AI chats as a governed data path, not just a productivity feature.
Q: What are the signs that legacy content controls are missing risky AI prompts?
A: The clearest sign is a mix of high false positives on harmless strings and silent misses on free-form text that clearly contains regulated or confidential material.
Practitioner guidance
- Implement semantic inspection for AI conversations Add context-aware detection to the conversation path so sensitive content is evaluated by meaning, not just by fixed patterns such as account numbers or ID formats.
- Separate structured and unstructured DLP policies Keep format-based controls for stable identifiers, but create a distinct policy layer for free-form prompts, replies, and mixed business text.
- Test policy quality with ambiguous prompts Use samples that contain overlapping meanings, such as names, locations, contract language, and medical references, to measure false positives and misses.
Bottom line: AI conversations can hide sensitive information inside ordinary prose, which means format-based DLP leaves a governance gap that semantic inspection is designed to close.
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
17/08/2026 4:43 pm
Context-aware detection is becoming a governance requirement, not a niche detection upgrade. AI conversation data does not behave like classic structured records, so controls that depend on fixed formats will increasingly underperform. Security teams need policies that evaluate meaning at the point of inference, because conversational systems now carry regulated content, secrets, and identity-linked data in ordinary prose. Practitioners should reframe AI data protection around semantic control.
A few things that frame the scale:
- 43% of security professionals are concerned about AI systems learning and reproducing sensitive information patterns from codebases, according to the State of Secrets in AppSec.
A question worth separating out:
Q: When should organisations combine DLP with identity and access governance for AI?
A: They should do it whenever AI tools can receive, summarise, or transform data that would normally be protected by classification or access rules. The control boundary moves to the prompt and response layer, so identity, authorisation, and content policy have to work together. Without that link, sensitive information can pass through a permitted user session into an uncontrolled AI workflow.
👉 Read our full editorial: Context-aware AI data detection closes the gap DLP still misses