Hello!

In the previous article, we walked through the mechanics and limits of HTTPS interception as the technical foundation of data-loss prevention in the AI era. From decrypting SSL/TLS traffic with a proxy server and operating an intermediate CA certificate to the technical constraint of certificate pinning, we laid out the technical realities enterprises face when monitoring web traffic.

These proxy techniques, however, are in fact common methods already widely adopted by existing DLP products. In traditional data-loss prevention — email monitoring, file-transfer control, web-access auditing — HTTPS interception has long played a central role.

So why does the AI era require us to rethink DLP anew?

Continuing from last time, this article focuses on the AI-specific challenges that traditional DLP cannot handle and the new considerations that AI-DLP must take on, developing the discussion at a more fundamental level.

1. The New DLP Requirements the AI Era Demands

Traditional DLP products have an excellent track record at detecting fixed patterns such as credit card numbers and social security numbers. These techniques will continue to serve as an important foundation. In the AI era, however, entirely new requirements have emerged on top of them.

Confidential information expressed in natural language

The biggest change is that handling confidential information expressed in natural language has become essential. Consider, for example, a query like
"We have a new product in development called Project Phoenix — could you propose an effective marketing strategy for it under that name?"
Handling this requires countermeasures that evolve beyond traditional keyword matching: a more sophisticated approach that recognizes "Project Phoenix" as a confidential codename and judges the information-leakage risk from context.

What makes this even more complex is that the same information carries different risk levels depending on context. The statement "our revenue is 120 billion yen" is unproblematic if it is a published prior-year figure, but a serious piece of confidential information if it is an unpublished current-year forecast.

There is also information that may be permitted for an internal AI tool but should never be sent to an external ChatGPT. This kind of context-dependent risk assessment requires a sophisticated mechanism that weighs multiple factors together: the organization's disclosure status, the type of destination AI service, the user's authorization level, and more.

As a conversation progresses, fragments of information combine into a highly confidential whole

Furthermore, the multi-turn nature of LLM dialogue creates a new class of challenge that traditional DLP never anticipated. Even if no single message contains confidential information, as the conversation proceeds, fragmentary pieces can combine to form a highly confidential overall picture.

Suppose, for example, an engineer starts with a generic technical question: "How can I compress large volumes of data efficiently in Python?" Then, discussing the algorithm ChatGPT proposed, they add concrete requirements —
"actually, the system we're building needs to process 10 GB of sensor data per second" — and continue with "our current proprietary compression ratio is 85%; is there a way to beat that?"
As the conversation goes on, they may unintentionally reveal their company's technical advantages and the details of a system under development.

It happens in human conversation as well: as the discussion heats up, you find yourself on the verge of accidentally letting confidential information slip.

With AI, people may in some cases confide even more freely than they would with another person.

Detecting this kind of cumulative, conversation-borne leakage risk requires deep AI-driven insight and the ability to continuously analyze the context of the conversation as a whole.

Even when each individual utterance looks harmless, the leakage risk that emerges when they combine is exactly where an AI-era DLP — an AI-DLP — proves its worth, and our LLM-Audit series works hard on precisely this.

The real-time nature of AI dialogue is another important constraint. For email or file transfer, a few seconds of pre-send inspection is tolerable; in a conversation with an AI, even a one-second delay degrades the user experience.

With email, sending the message concludes a unit of work, so a slower, audited send builds little stress. A conversation with an AI, however, proceeds in real time just like a conversation with a person — if an AI-DLP audit makes the user wait several seconds at every turn, the reaction becomes "these replies are too slow; I can't work like this."

Within this tight time budget, we need technology that maintains the context of an ongoing conversation and can instantly evaluate how risk shifts across a single utterance or a multi-turn series of utterances.

2. Bidirectional Auditing: a New Paradigm

Traditional DLP focused on preventing information from flowing out of the organization — that is, on outbound protection. The risks of sending confidential information, leaking personal data, and losing intellectual property certainly remain important. In the AI era, however, that alone is not enough.

Responses from the AI — inbound information — are becoming just as important, if not more so. AI-generated output can contain misinformation and hallucinations, inappropriate content, advice that violates compliance, and even malicious code. These bring new risks to the organization.

For example, what happens when a legal-department employee consults an AI about contract clauses and the AI provides a mistaken legal interpretation? Or when an engineer asks an AI to optimize code and the generated code contains a security hole? These risks were simply never contemplated by traditional DLP.

Managing the quality of AI responses has become critical to preserving the quality of organizational decision-making. It requires multifaceted verification: factual accuracy — especially checking figures and dates — legal and ethical appropriateness, consistency with organizational policy, and detection of latent bias. Technically this is very demanding, but it is more than a technical challenge: it reaches into organizational governance itself, which makes it a key point of discussion.

3. Technical Approaches Specific to AI-DLP

Meeting these new requirements calls for technical approaches that differ from the traditional ones. At their core is a tiered processing architecture.

Inspecting all traffic in equal detail is unrealistic from a performance standpoint. Instead, an effective approach dynamically adjusts the depth of processing according to risk level: fast screening quickly passes clearly unproblematic traffic, and only suspicious traffic is sent to deeper analysis.

At the first stage, basic keyword matching and known-pattern detection run within 5 milliseconds. That alone should let more than 95% of all traffic pass without issue. The next stage performs more detailed structural analysis using natural-language-processing techniques such as morphological analysis and named entity recognition (NER). With user experience in mind, this too must complete on the order of 100 milliseconds.

When still more advanced judgment is needed, contextual analysis is performed with deep-learning models such as BERT, DeBERTa, and RoBERTa. These models understand the meaning of text deeply and can capture subtle nuance. They are also well suited to assessing the risk of text accumulated across multi-turn conversations, as in the example at the start of this article.
Finally, in the most complex cases, a purpose-trained LLM performs a comprehensive judgment, taking into account organization-specific business rules and past decisions to reach the final allow/deny verdict.

The advantage of this tiered approach is that the vast majority of harmless traffic incurs minimal delay, while genuinely risky traffic receives ample analytical resources. This achieves the optimal balance between security and usability.

4. The Importance of Continuous Learning and Adaptation

An AI-DLP system is not a static, rule-based system. It must learn the organization's specific context and improve its accuracy over time.

Let us bring back "Project Phoenix" from our opening example.

Suppose an organization begins using the codename "Project Phoenix." Initially, the system does not know this is a confidential project. But by detecting that the term appears frequently in confidential documents and restricted communications, it gradually learns its significance. In the same way, new product names, department names, and technical terms are automatically added to the dictionary and assigned appropriate confidentiality levels.

Feedback from false positives and misses is another important source of learning. When a security officer marks an alert as a false positive, the system learns that decision pattern and reduces similar false positives going forward. Conversely, when a missed leak of confidential information comes to light later, the system learns that pattern too and improves its detection capability.

Through this continuous learning, an AI-DLP system becomes able to make increasingly accurate judgments grounded in the organization's realities. This is an evolution unique to the AI era — something traditional DLP, dependent on fixed rules, could never achieve.

5. The Trinity of Technology, Usability, and Governance

The discussion so far has centered on technology, but technology alone is not enough for AI-DLP to succeed. As we noted in the previous article, 100% technical defense is impossible, and in the end human awareness and ethics matter.

An overly restrictive security policy can actually undermine an organization's security. Ban AI tools outright, or make them usable only in extremely inconvenient ways, and employees will start accessing these tools from personal devices and personal accounts.

This state of affairs —

"the company network is such a pain, I'll just quietly do things my own way"

— is called "shadow IT" (Shadow IT).

Once shadow IT takes hold, it completely spoils organizational governance.

What matters instead is a shift in mindset from "prevent use" to "enable safe use" — capturing AI's productivity benefits for employees while managing the risks appropriately. That is the essence of security in the AI era.

On the technical side, the tiered processing described in this article leaves the vast majority of normal use unaffected and inspects only genuinely high-risk traffic in detail. On the organizational side, establish clear usage guidelines and graduated restrictions by risk level — relatively light restrictions for general employees, stricter ones for departments handling confidential information. Flexible operation is key.

Regular awareness programs and education are also indispensable. Why are these measures necessary, what are the risks, and how can AI be used safely? Genuine security improvement comes when employees understand and accept these answers and put them into practice.

6. The Future of AI-DLP

In thinking about the future of AI-DLP, it is important to plan on short-, medium-, and long-term horizons.

In the short term, the priority is to make the most of currently available technology and build a basic defensive posture. Combining the proxy-based monitoring covered last time with the tiered processing described in this article achieves a practical level of security.

In the medium term, new solutions specialized for AI-DLP will continue to evolve. The accuracy of natural-language understanding and context awareness will improve, enabling more sophisticated judgments. Control methods optimized for the characteristics of each different AI service should also become established.

In the long term, we will see the era of the "guardian agent" — using AI technology itself, comprehensively, to manage the risks AI creates. Advanced language models will judge the confidentiality of data about to be sent with accuracy equal to or better than a human's and apply the appropriate controls. That future should not be far off.

7. Toward Practical Solutions

This article has examined the new challenges in AI-era data-loss prevention that traditional DLP approaches cannot address, and the technical approaches to meet them. The challenges AI-DLP must tackle are wide-ranging: the need for natural-language understanding, context-dependent risk assessment, real-time processing, and the new paradigm of bidirectional auditing.

We have also seen that technical solutions such as tiered processing architectures and continuous learning mechanisms are effective against these challenges. At the same time, much cannot be solved by technology alone, and organization-wide effort remains indispensable.

LLM-Audit: a Comprehensive AI-DLP Solution

Our LLM-Audit can serve as a comprehensive AI-DLP solution, developed with a deep understanding of these challenges. LLM-Audit is designed not as a mere technical product but as an integrated platform for advancing an organization's AI adoption safely.

LLM-Audit's defining feature is its implementation of the bidirectional auditing described in detail in this article. Outbound auditing inspects the data employees are about to send to an LLM, checking for personal or confidential information. Inbound auditing, in turn, inspects the LLM's responses for inappropriate content, misinformation, and unintended exposure of confidential information.

Thanks to its flexible tiered-guardrail architecture, more than 95% of ordinary requests are handled by fast matching (around 5 milliseconds), so users perceive virtually no delay. For more complex cases, processing deepens in stages as needed: morphological analysis, named entity recognition, contextual analysis with BERT/RoBERTa, and inference by a dedicated LLM.

We also offer PII-FI, which performs only PII and confidential-information detection and pseudonymization (de-identification) at high speed and accuracy — and we will propose the solution best suited to your use case.

Conclusion

Data-loss prevention in the AI era is an entirely new challenge that can no longer be met as an extension of the old approaches. On top of the foundational technology of HTTPS interception, we must layer new elements: natural-language understanding, context-dependent judgment, real-time processing, and bidirectional auditing.

Yet technical evolution alone is not sufficient. The shift in mindset from "prevent use" to "enable safe use," an approach uniting technology, usability, and governance, and an established cycle of continuous improvement — these are the keys to achieving genuine security in the AI era.

Seek not perfection but practical, sustainable measures. Recognize the limits of technology, and press forward regardless. That is the posture the AI era demands of security.

We believe that transparency of technology is the first foundation of security. We hope this article helps inform your organization's discussions on reconciling AI adoption with security.

As AI technology evolves, security challenges shift by the day.

If your company is grappling with security and governance challenges in AI adoption, please do reach out. Building on the technical background covered in this and the previous article, we will propose the solution best suited to your specific situation.