Secure Service Edge · Topic 4 of 5
Data Loss Prevention and AI Guardrails for Secure Internet Access
Learn how DLP policies and AI guardrails in Cisco Secure Access stop sensitive data leaking through web, SaaS and generative AI applications.
- 5 min
- Real time DLP inspects traffic in line. API DLP scans data at rest in SaaS applications.
- Data classifications use built in or custom patterns, keywords and context.
- Start DLP in monitor mode, tune it, and then block.
What it is
Data loss prevention (DLP) detects sensitive data that is about to leave, and AI guardrails apply similar control to generative AI use.
Users upload files, paste text into web forms and send prompts to AI tools. Any of these can leak customer records, source code or credentials. In Cisco Secure Access, the cloud service inspects this traffic and enforces policy.
How it works
Define what sensitive data looks like, choose where to look, and decide what action to take.
Where DLP can look
DLP works at two points, and they give different coverage.
| Mode | What it inspects | Typical action |
|---|---|---|
| Real time (in line) | Web and SaaS traffic as the user sends it | Block or monitor an upload or a post |
| SaaS API (at rest) | Data already stored in connected SaaS applications | Report, or remediate where the service allows it |
Real time DLP works on traffic that passes through the service, so it needs the traffic to be steered there, and needs TLS decryption for encrypted sessions. API based DLP connects to a SaaS tenant and scans files that are already there, including files shared outside the company.
Data classifications
A data classification describes the kind of data to find. The service includes built in classifications for common data types, and you can create your own.
- Built in: common patterns such as payment card numbers and personal identifiers, grouped by regulation or region.
- Custom: your own patterns, regular expressions or keyword dictionaries.
- Proximity and context: a keyword near a number lowers false positives.
A pattern alone causes false alarms. A sixteen digit number may not be a card. Validation, such as a checksum, and nearby keywords improve accuracy.
Building a DLP policy
A DLP policy ties the classifications to the traffic and the action.
| Part | Meaning |
|---|---|
| Identities | Users or groups that the policy covers |
| Destinations | Applications, categories or sites, such as file sharing or webmail |
| Data classifications | What counts as sensitive |
| Direction | Usually uploads and posts |
| Action | Monitor, or block with a notification |
Start in monitor mode. Review the matches, remove noise, and tune thresholds. Then move to blocking for the high confidence cases.
Handling encrypted traffic and files
Inspection needs readable content, so a DLP rule that matches HTTPS needs the traffic to be decrypted. Files are scanned for text, including inside common document types and archives. Very large or unsupported files may be skipped or handled by a stated policy, so know the limits.
AI guardrails
Generative AI tools create a new path for data to leave, and a new path for attacks. Guardrails control that use. They have two sides.
- Access control: decide which AI applications are allowed, monitored or blocked, and for whom.
- Content inspection: check the prompts that users send, and the responses that come back.
On the prompt side, inspection can find sensitive data, such as customer records or secrets, before they reach an external model. It can also detect prompt injection attempts. On the response side, it can detect harmful or policy breaking content.
| Risk | Guardrail response |
|---|---|
| Staff paste confidential data into a public AI tool | Block the prompt or warn the user |
| Use of unapproved AI applications | Block or restrict the application by category |
| Prompt injection | Detect and block suspicious input |
| Harmful or off policy responses | Detect and block the response |
The product capabilities and names for AI controls are changing quickly, so check the current documentation for the exact options in your release.
Operating the controls
Tell users what is blocked and why, because a clear notice reduces help desk calls. Keep an exception process for legitimate needs. Review reports regularly for repeated attempts, which point to training needs or to a real insider risk.
What to remember for the exam
Six facts cover most questions on this topic.
- Real time DLP inspects traffic in line. API DLP scans data at rest in SaaS applications.
- Data classifications use built in or custom patterns, keywords and context.
- Start DLP in monitor mode, tune it, and then block.
- DLP on HTTPS needs TLS decryption.
- AI guardrails control which AI apps are used, and inspect prompts and responses.
- Guardrails address data leakage and prompt injection.
If a question asks how to find sensitive files already sitting in a SaaS application, choose API based DLP.
Common mistakes
These four mistakes appear often.
Blocking at once without tuning. False positives disrupt legitimate work.
Forgetting decryption. The DLP engine cannot see encrypted content.
Relying on real time DLP for stored data. In line inspection sees only traffic that passes through it.
Banning AI with no alternative. Users move to unmanaged tools, so offer an approved option with guardrails.
DLP reduces accidental loss. It does not stop a determined insider who can move data by another route.
Check yourself
Work through these three questions.
What is the difference between real time DLP and API based DLP?
Real time DLP inspects traffic as it passes through the service, and API DLP scans data that is already stored in a connected SaaS application.
Why start a DLP policy in monitor mode?
To review matches and tune out false positives before blocking real work.
What two things do AI guardrails control?
Which AI applications may be used, and the content of prompts and responses.
Topic 4 of 5 in Secure Service Edge
Test yourself on Secure Service Edge
The full Cisco CCNP 350-701 SCOR question bank has 822 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Cisco. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.