The Great AI Memory Bank: How Your Data Gets Consumed (and How to Keep It Private)

In a nutshell (TL;DR)

To secure your data, consider these strategies:

• Anonymization Pipelines: Replace sensitive identifiers with placeholders (e.g., [NAME]) before data leaves your network.
• Zero Data Retention (ZDR): Mandate that providers process prompts in memory only, without saving logs or using data for training.
• Local Models & Secure Orchestration: Keep data within corporate firewalls by running local models or utilizing secure protocols like MCP.
• Targeted Encryption: Encrypt or mask sensitive prompt segments, such as using unique emoji sequences, to keep text unreadable to the provider.

It's been two weeks since I last posted! But I am back after the day job got in the way with a major project and a tight deadline. Last post I talked about the dangers of copy and paste and how easily information can end up in the hands of the LLMs

Whenever we type a prompt into an AI assistant, it is easy to imagine our words vanishing into the digital ether the moment we hit 'send'. But Large Language Models (LLMs) have incredibly sticky memories. While it is easy to accidentally slip sensitive data into an AI tool, it is equally important to understand what the AI actually *does* with that information once it has it. 

LLMs are designed to consume, process, and generate text, which means treating them like a private diary or a secure vault can lead to unintended, and highly public, consequences. Here is a look at how your confidential information gets consumed and redistributed by AI, and the best practices you can use to keep your private data safe.

The Consumption and Redistribution Cycle

When you feed Personally Identifiable Information (PII) or corporate secrets into an external LLM, you are exposing that data to several hidden risks:

Data Logging and Storage

Many AI providers log user prompts to monitor for abuse, debug their systems, or improve their overall services. Once your confidential data is stored on a third-party server, it becomes vulnerable to unauthorized access or potential data breaches on the provider's end.

Training Data Contamination

The prompt you submit today could inadvertently become the training data of tomorrow. Even though some enterprise providers have strict policies, there is always a baseline risk that PII from user prompts might be absorbed to further train or fine-tune future versions of the models.

Output Leakage and Regurgitation

LLMs are known to memorize information from their pre-training phases as well as from prompts processed during active inference. This can lead to a phenomenon where the model unintentionally regurgitates your sensitive information verbatim in its responses to completely different users. In fact, the OWASP Top 10 for LLMs lists "Sensitive Information Disclosure" as a critical vulnerability, noting that poor input handling can cause models to leak PII, business strategies, or system credentials directly into the public domain.

Defending Your Data: Precautions and Safe Methods

Fortunately, you do not have to unplug your routers and swear off AI entirely. There are several highly effective precautions and architectural strategies you can implement to interact with LLMs safely:

1. Build Anonymization and Mapping Pipelines

The most practical defense is to scrub the data before it ever leaves your network. By using tools like Named Entity Recognition (NER), you can automatically identify sensitive entities and replace them with generic placeholders—for example, swapping a real name and email for `[FIRSTNAME]` and `[EMAIL]`. This allows the LLM to understand the context of the prompt without ever seeing the raw data. On your end, you keep a secure, temporary map of these placeholders. When the LLM replies, a mapping-based de-anonymization module simply swaps the real information back in, ensuring 100% accuracy without exposing the data to the cloud.

2. Demand Zero Data Retention (ZDR)

If you rely on cloud-based AI vendors, mandate a "Zero Data Retention" agreement. Under ZDR, the provider processes your prompt and immediately returns the response without writing your request to any persistent storage, training queues, or logs. The data exists only in memory for the exact duration of the API call, effectively shifting your risk profile from uncertain to bounded.

3. Utilize Local Models and Secure Orchestration (e.g., MCP Servers)

For the highest level of control, organizations can run fine-tuned, smaller language models entirely within their own corporate firewalls, ensuring data never leaves the internal infrastructure. When connecting AI to internal databases, utilizing secure architectural patterns like the Model Context Protocol (MCP) can help safely orchestrate how context is provided to the AI without exposing raw data to public endpoints.

4. Targeted Encryption

For highly regulated environments, researchers are developing targeted encryption techniques. This involves encrypting only the sensitive sub-parts of a prompt, sometimes even translating them into unique sequences of emojis (like *EmojiCrypt*), so the text remains unreadable to humans and providers, but retains enough structure for the LLM to process. While computationally expensive and complex to implement, it represents the bleeding edge of prompt privacyLarge Language Models (LLMs) pose significant security risks because they can unintentionally memorize and redistribute sensitive information, such as PII and corporate secrets. Primary dangers include unauthorized data logging, training data contamination, and output leakage where models regurgitate your data to others.

AI models are incredibly eager to learn, which makes them fantastic assistants but terrible secret-keepers. By adopting smart anonymization pipelines, demanding strict retention policies, and securing your integrations, you can enjoy all the productivity benefits of generative AI without accidentally donating your private data to the world.

The Clipboard Crisis: Securing the Modern Data Leakage Vector

In a nutshell (tl;dr)

The modern copy-paste function has become a major, often overlooked, vector for data exfiltration. As employees frequently use unmanaged personal accounts for Generative AI and messaging apps, corporate data is regularly moved outside secure environments. Because traditional security tools were designed to monitor file uploads rather than "file-less" text transfers, organizations must shift toward action-centric security, monitor browser activity, and restrict the use of personal accounts to protect sensitive information.

I totally missed out on last week’s post thanks to the day job and nearly missed it this week too! This week I was worried about how safe my personal or proprietary data was when passing it over to an AI to work with. How likely is it that this information is somehow leaked or made public? Here’s what I found out…

The Copy-Paste Crisis

We all use the copy-paste function without a second thought and the clipboard is our biggest blind spot. It is the ultimate productivity shortcut, saving us countless hours of retyping information. However, this simple, everyday action has quietly become the primary channel for data exfiltration in the modern workplace, completely bypassing traditional file-based security measures.

As we increasingly rely on artificial intelligence and cloud-based applications, the clipboard has transformed into a massive vulnerability. Here is a detailed look at how the "copy-paste crisis" unfolds, why it is so dangerous, and what organizations can do to protect their confidential data.

The Generative AI Black Hole

Generative AI tools have seamlessly integrated into our daily routines, and we are eagerly feeding them information to summarize, rewrite, or analyze. In fact, a staggering 77% of enterprise employees now paste data directly into GenAI prompts.

The core issue is not necessarily the AI itself, but how users are accessing it. Approximately 82% of the data pasted into AI tools comes from unmanaged, personal accounts. When employees bypass official corporate logins, IT and security departments lose all visibility. This turns "Shadow AI" activity into a massive blind spot for data leakage. Today, GenAI alone accounts for 32% of all corporate-to-personal data exfiltration, making it the number one vector for corporate data moving outside sanctioned environments.

Beyond AI: The Instant Messaging Trap

While AI gets most of the spotlight, instant messaging (IM) and chat applications represent another enormous vulnerability. A remarkable 87% of all instant messaging activity occurs on unmanaged, non-corporate accounts .

Because chat feels informal and conversational, users often let their guard down. Consequently, Chat and IM apps have become a major hotspot for sensitive data exposure, with 62% of users pasting Personally Identifiable Information (PII) or Payment Card Industry (PCI) data directly into these platforms.

Death by a Thousand Clicks

It might seem like pasting a quick snippet of text is harmless, but the sheer volume of these actions adds up to a significant security threat. On an average day, an employee makes about 46 copy-paste actions. While many of these transfers stay safely within corporate boundaries, an average of 15 pastes per day go to non-corporate accounts. Out of those, roughly four pastes contain sensitive PII or PCI data.

An employee pasting a few sensitive entries into ChatGPT each day might not trigger massive security alarms or generate large file logs, but every single instance increases the risk of a breach. Furthermore, employees are pasting corporate data into a surprisingly diverse range of destinations. Beyond just ChatGPT, top destinations for pasted data include developer platforms like Databricks and Snowflake, as well as websites like LinkedIn and DeepL. Exfiltration is highly unpredictable, driven by everything from innocent productivity shortcuts to competitive moves.

Why Traditional Defenses Are Falling Behind

The reason this copy-paste crisis has grown so severe is that traditional Data Loss Prevention (DLP) solutions were fundamentally designed for a different era. Legacy DLP focuses heavily on monitoring file uploads and centralized servers . They simply are not equipped to track "file-less" data transfers, like copying text from an internal document and pasting it directly into a web browser.

Taking Back Control of the Clipboard

To secure the modern workflow, organizations need to evolve their security strategies to match employee behavior.

Shift to Action-Centric Security

Security teams must move away from purely file-centric policies and embrace "action-centric" controls. Monitoring copy-paste functions and text inputs into prompts must become a first-class security priority.

Focus on the Browser

Because nearly every business workflow, from email to GenAI, now runs through the web browser, this is the environment where visibility and enforcement must be focused.

Ban Unmanaged Accounts

Allowing employees to use personal accounts for business-critical apps creates active shadow IT. Organizations should restrict the use of personal accounts for high-risk categories like AI and Chat, and enforce Single Sign-On (SSO) across all corporate logins to ensure activity remains visible and governed.

The clipboard might be the most overlooked tool in our software arsenal, but it is currently one of the riskiest. By understanding the flow of copy-pasted data and upgrading our security frameworks to monitor file-less transfers, we can enjoy the productivity benefits of modern SaaS and AI tools while keeping our private data exactly where it belongs.

Fortifying the Digital Vault: A Wee Guide to AI Privacy

In a nutshell (TL;DR)...

The widespread use of generative AI tools introduces major security risks for private and confidential company information. Sensitive data can leak when prompts are retained for logging/training, employees paste data into unmanaged "Shadow AI" accounts (the "Copy/Paste Blind Spot"), or malicious "Prompt Injections" trick the model. Consequences are severe, including regulatory fines (GDPR/HIPAA), data breaches, and loss of competitive advantage. To stay secure, organizations must:

• Anonymize sensitive data (PII) before using external LLMs.

• Prioritize vendors offering Zero Data Retention (ZDR).

• Banish "Shadow AI" by enforcing Single Sign-On (SSO).

• Upgrade to action-centric Data Loss Prevention (DLP) that monitors copy/paste actions.

Apply the principle of least privilege and keep a human in the loop for critical actions.

The AI Privacy Guide: How to Keep Your Confidential Data Safe in the Age of LLMs

The company I work for has drummed into me the perils of letting slip any confidential information when working with AI applications, but just how important is it? My employer specifically lists the AI applications we are allowed to use when working with confidential information, so it’s a really important thing to bear in mind. Let’s have a look at what the problems are and how we can protect ourselves, our customers and our employers…

Everyone is officially living in the era of Artificial Intelligence. From drafting emails to analyzing complex datasets, generative AI and Large Language Models (LLMs) have seamlessly integrated into our daily workflows. In fact, nearly half of all enterprise employees are already using these tools. But amid all this newfound productivity, there is a crucial conversation we need to have: how are we protecting our private data and confidential company information?

While AI assistants are incredibly helpful, treating them like a private diary or a secure company vault can lead to serious risks. Let’s break down exactly how sensitive information can slip through the cracks, what the consequences are, and the best practices you should adopt to stay secure.

How Does Confidential Information Actually Go Public?

When you type a prompt into an external LLM, that data is processed by a third-party provider. If you aren't careful, sensitive information can be exposed in a few common ways:

Logging and Training Contamination

Many AI providers retain user prompts for a certain period to monitor for abuse, debug their systems, or even train future versions of their models. If you paste confidential data into a prompt, it could end up stored on the provider's servers or, worse, replicated in the model's future outputs.

The Copy/Paste Blind Spot

A staggering 77% of employees paste data directly into generative AI tools, and the vast majority of this activity happens on unmanaged personal accounts. Because this bypasses official corporate channels, IT and security teams have no visibility into what is being shared, creating a massive "Shadow AI" blind spot.

Prompt Injections

Malicious actors can use "prompt injections", carefully crafted inputs designed to manipulate the AI's behavior to trick the model into revealing sensitive information. This can lead to the AI accidentally exposing personally identifiable information (PII), confidential business strategies, or even system credentials. I’ve made a note to dig deeper on this subject for a later post…

The Uncomfortable Consequences of Data Leaks

The fallout from exposing sensitive data to an LLM is rarely a minor hiccup. When PII or corporate secrets leak, the consequences can be severe.

Regulatory Penalties

Mishandling personal data violates strict data protection regulations like GDPR and HIPAA. Failing to comply with these laws can result in massive legal and financial penalties.

Data Breaches and Loss of Trust

If a customer service chatbot or an internal AI tool inadvertently reveals private user details or passwords, it can lead to full-scale data breaches. This erodes user trust and severely damages your organization's reputation.

Loss of Competitive Advantage

Exposing proprietary business data or intellectual property can directly result in a loss of your competitive edge in the market.

Best Practices for Handling Sensitive Information with AI

Fortunately, you don't have to abandon AI to keep your data safe. By implementing a few strategic best practices, you can enjoy the benefits of LLMs while minimizing your risk.

1. Anonymize Before You Analyze

Before sending a prompt containing sensitive data to an external LLM, scrub the text of any PII. You can use automated tools to detect and replace names, emails, and phone numbers with generic placeholders (e.g., swapping a real name for [PERSON] or [EMAIL]). This allows the AI to understand the context of your prompt without ever seeing the raw, sensitive data.

2. Demand "Zero Data Retention" (ZDR)

If you are procuring AI tools for your company, prioritize vendors that offer "Zero Data Retention" agreements. Under a ZDR policy, the AI provider processes your prompt and immediately returns the response without writing your data to any persistent storage, logs, or training queues. This ensures your data exists only in memory for the duration of the request. I think this is what my employer might have in place for the applications I am allowed to use.

3. Banish "Shadow AI" and Enforce SSO

Employees often use unmanaged personal accounts to access AI tools, completely bypassing enterprise security. To regain control, organizations must restrict the use of personal accounts for business-critical apps and enforce Single Sign-On (SSO) across all corporate logins.

4. Upgrade Your Data Loss Prevention (DLP)

Traditional Data Loss Prevention tools are heavily focused on file uploads, but today's sensitive data usually leaks when employees copy and paste text directly into AI prompts. Organizations need to shift to "action-centric" DLP policies that monitor file-less data transfers and enforce controls directly at the web browser level.

5. Keep a Human in the Loop and Limit Privileges

Finally, never give an AI unchecked autonomy. Apply the principle of "least privilege" by ensuring your AI applications only have access to the specific data sources they absolutely need. For high-impact actions, like modifying files or handling highly sensitive records, always require human approval before the AI can proceed.

AI is a powerful collaborator, but it is ultimately up to us to set the boundaries. By treating generative AI platforms with the same security rigor as any other enterprise tool, we can innovate quickly without putting our most valuable data on the line.

Next week let’s take a shifty at this “prompt injection” malarky and see how we can protect ourselves from that…

A Comprehensive Guide to the 2026 LLM Landscape

Architecting Logic: The Chain-of-Thought Prompting Guide

Building on our exploration of Zero-Shot and Few-Shot techniques, Chain-of-Thought (CoT) prompting is the natural next step for tackling tasks that require deep logic. While standard few-shot prompting is excellent at teaching a model what format to output, it often fails at teaching the model how to process a complex problem.

Here is a detailed overview of Chain-of-Thought prompting, why it is critical, and how to implement it effectively.

What is Chain-of-Thought Prompting?

Chain-of-Thought prompting is a technique designed to improve the reasoning capabilities of Large Language Models (LLMs) by forcing them to generate intermediate reasoning steps before providing a final answer. Instead of treating the AI as a "black box" that spits out an immediate conclusion, CoT guides the model to mimic human problem-solving by "talking through" the problem step-by-step.

Why is it Important?

CoT is vital for several reasons:

• Overcoming Few-Shot Limitations: Standard few-shot prompting provides the final answer but skips the underlying logic, meaning the model fails to learn the process. CoT bridges this gap, allowing models to succeed on complex arithmetic, common-sense, and symbolic reasoning tasks.

• High Impact, Low Effort: It is a highly effective way to boost performance using off-the-shelf LLMs without the need for expensive fine-tuning.

• Interpretability: Because the model outputs its reasoning, you gain transparency into its "thought process." If the model hallucinates or makes a mistake, you can read the steps to identify exactly where the logic broke down.

• Robustness: CoT prompts tend to experience less performance drift when migrating between different LLM versions compared to standard prompts.

Solid Examples of CoT in Action

To understand CoT, let's look at a classic mathematical failure and how CoT fixes it.

Standard Prompt (The Failure):

Prompt : When I was 3 years old, my partner was 3 times my age. Now, I am 20 years old. How old is my partner?

Output: 63 years old.

(The model blindly multiplies without reasoning through the age gap).

1. Zero-Shot Chain-of-Thought

You can trigger CoT without providing any examples simply by appending a specific instruction, most commonly "Let's think step by step".

Prompt : When I was 3 years old, my partner was 3 times my age. Now, I am 20 years old. How old is my partner? Let's think step by step.

Output :

1. When I was 3 years old, my partner was 3 * 3 = 9 years old.

2. That is an age difference of 6 years.

3. Now I am 20 years old, so my partner is 20 + 6 = 26 years old.

4. The answer is 26.

2. Few-Shot Chain-of-Thought

For even better results, you can combine CoT with few-shot learning by providing examples where the reasoning process itself is mapped out in the demonstration.

Prompt :

Q: When my brother was 2 years old, I was double his age. Now I am 40 years old. How old is my brother? Let's think step by step.

A: When my brother was 2, I was 2 * 2 = 4. That's an age difference of 2 years. Now I am 40, so my brother is 40 - 2 = 38.

Q: When I was 3 years old, my partner was 3 times my age. Now, I am 20 years old. How old is my partner? Let's think step by step.

A: 10

Use Cases for Implementation

Generally, any task that a human would solve by "talking it through" is a great candidate for CoT.

Specific use cases include:

• Mathematical and Logical Reasoning: Solving complex word problems, physics questions, or symbolic logic puzzles where jumping straight to the answer causes hallucinations.

• Code Generation and Debugging: Breaking a software request down into functional steps before mapping those steps to specific lines of code.

• Synthetic Data Generation: Guiding a model to systematically think through the assumptions and target audience of a product before writing a description for it.

Are There Downsides to This Technique?

While powerful, CoT is not a silver bullet and comes with several notable downsides:

Increased Cost and Latency

Because the model must generate the intermediate reasoning text before delivering the final answer, it consumes significantly more output tokens. This means your predictions will cost more money and take longer to generate.

Strict Temperature Requirements

CoT relies on "greedy decoding"—predicting the most logically probable next word. To use CoT effectively, you must set the model's temperature to 0 (or very low), which limits its use in creative tasks.

Diminishing Returns on the Newest Models

Recent research indicates that highly advanced foundation models (like Qwen2.5 or DeepSeek-R1) have been exposed to so much CoT data during training that they have internalized these reasoning patterns. For these extremely strong models, adding traditional CoT exemplars often fails to improve reasoning ability beyond standard zero-shot prompting, as the models simply ignore the examples and rely on their internal knowledge.

API Policy Restrictions

For newer, dedicated "reasoning models" (like OpenAI's o-series), the models handle the chain of thought internally. Attempting to manually extract or force CoT reasoning through prompts is often unsupported and can even violate Acceptable Use Policies 14.

Chain-of-thought (CoT) prompting remains a cornerstone of LLM interaction, but its role has shifted from a "magic trick" that fixes everything to a specialized tool that must be used strategically.

Relevancy and usefulness of CoT today

In the current landscape of 2026, the relevance of CoT depends entirely on whether you are using a Reasoning Model (like OpenAI’s o1/o3 or Gemini Flash 2.5) or a Standard Model (like GPT-4o or Claude 3.5 Sonnet).

1. For Standard Models (GPT-4o, Claude 3.5, Gemini 1.5 Pro)

CoT is still highly useful but inconsistent. Recent studies show that "thinking step-by-step" provides a significant boost on complex logic, but can actually degrade performance on simple tasks.

• The "Thinking" Tax: Using CoT increases latency by 35% to 600% and scales token costs proportionally.

• Performance Gains: Models like Claude 3.5 Sonnet still see accuracy improvements of roughly 10–12% on complex reasoning tasks when prompted with CoT.

• The Inconsistency Risk: Paradoxically, Gemini 1.5 Pro and GPT-4o sometimes perform worse (-17% in some benchmarks) when forced to use CoT on "easy" questions they would have otherwise answered correctly via intuition.

2. For Reasoning Models (OpenAI o1/o3, Gemini 2.0/2.5)

CoT prompting is becoming redundant. These models have "Internal CoT" baked into their architecture—they reason before they speak by default.

• Diminishing Returns: Explicitly adding "think step by step" to a model that is already designed to think (like o3-mini) yields marginal gains (often <3%) while significantly increasing the time-to-first-token.

• Conflict of Logic: In some cases, forcing an external chain of thought can interfere with the model’s internal reinforcement-learned reasoning paths, leading to "overthinking" errors.

Comparison: When to Use CoT

Task Type	Utility	Recommended Approach
Simple Extraction/Q&A	❌ Harmful	Ask for a direct answer to save cost and avoid "hallucinated" logic.
Complex Math/Coding	✅ Critical	Use CoT or, better yet, use a dedicated Reasoning Model.
Creative Writing	⚠️ Mixed	CoT can make the output feel "formulaic" or robotic.
Policy/Compliance	✅ High	Use CoT for explainability—it’s useful for auditing why a model made a decision.

The Modern "Best Practice"

Instead of the generic "Let's think step by step," the current trend is Structured CoT. Rather than letting the model wander, you define the "steps" you want it to take:

1. Analyze the user's intent.

2. Identify relevant variables/constraints.

3. Draft a logical solution.

4. Verify the solution against the constraints.

Summary

CoT is not a universal "better" button anymore. It is a precision tool. If you are using the latest reasoning models, you can likely retire the "step-by-step" prompt entirely. If you are using standard models for complex logic, it remains your best defense against "hallucinated" shortcuts—just be prepared to pay for it in latency and tokens.