AI tools are becoming more and more common in the development workflow. With the boom of AI in recent years, we have seen an increasing number of tools available to help developers in their daily tasks. They can help us write test and code faster, find bugs, and in general, be more productive if we have good knowledge of what we are doing.
But there are risks.
In this post, we’ll take a look at some of the most popular AI tools for developers and discuss their capabilities, privacy concerns and learning concerns.
GitHub Copilot
GitHub Copilot is most likely the first AI tool a developer has introduced in their workflow. It’s available as a plugin for Visual Studio Code and other popular code editors. It supports a wide range of programming languages and frameworks, making it a versatile tool for developers working on different projects.
How it works
It uses popular models trained on a large dataset of code to generate suggestions that are relevant to the code we are writing. The models available as per March 2025 are OpenAI 4o, o1 and o3-mini, Claude 3.5, 3.7, and 3.7 thinking and Gemini 2.0 Flash.
To get the best out of GitHub Copilot I recommend having a look at their documentation as they have a cookbook section with examples on how to use it.
In my experience Copilot is great to help us be faster writing tests or refactoring, but it’s not that great when we are working in a complex production application. The majority of the time the suggestions are not very helpful even if you include the files in the chat so the model can have more context. This leads to distractions and lost time.
Privacy concerns
Like the majority of AI tools, Copilot requires access to our codebase to generate suggestions, which raises questions about data privacy and security. It is important to understand how Copilot uses our code and what data it collects to ensure that our code and data are secure.
GitHub Copilot sends information to GitHub’s Azure servers to generate code and chat suggestions. This information includes context about our code and files as well as data about how we interact with the service. All data is protected with industry-standard encryption during transfer and storage.
The service processes personal information differently depending on how we access it through the website, mobile app, IDE extensions, or when using features like CLI suggestions, code completions, or chat. The collected information falls into several categories: user interaction metrics (like which suggestions we accept or reject), our inputs for generating suggestions, the AI-generated responses themselves, and any feedback we provide about the service.
It’s important to note that GitHub has explicitly stated they do not use data from Copilot Business or Enterprise customers to train their AI models.
The fact that Business and Enterprise data isn’t used for model training suggests GitHub is sensitive to privacy concerns for professional users. In any case we should remain aware that our coding activity and interactions with Copilot are being monitored and stored to improve the service and generate personalized suggestions.
Cursor.ai
Cursor.ai is an AI-powered code editor with integrated AI assistance. It’s essentially a modified version of Visual Studio Code that incorporates advanced AI capabilities directly into the coding environment.
How it works
When we use Cursor the code and contextual information is processed by various AI models, including those from OpenAI, Anthropic, and Google, to generate suggestions and responses. The tool can analyze the current file, surrounding code context, and project structure to provide more relevant assistance.
What sets Cursor apart from Copilot is their integrated AI features. Code generation from natural language prompts, contextual chat assistance, and codebase understanding capabilities are natively integrated in the interface.
Cool features
Cursor’s Codebase Indexing feature creates semantic representations of our code to enable more contextual AI assistance. While enabled by default, this feature can be disabled during setup or in settings. When indexing is active, Cursor creates encrypted representations of our files, with paths obfuscated to protect sensitive information.
One cool thing with Cursor is that we can add some project rules. These rules
are a set of instructions for the model to know how it should code. The rules
live in .cursor/rules. We can find some in this cursor directory.
For my Rails personal projects I’ve been using thoughtbot guides
plus some of the directions from here.
In my experience, the context awareness, the indexing and the integrated chat assistance capable of analysing different files at the same time is a game changer. I have some knowledge of React Native and TypeScript, but I’m not an expert.
With the help of Cursor I was able to write a working mobile app in one day with some basic features. Design included. Obviously it is not a production ready application but to conceptualise an idea and have a working prototype in a day is amazing.
Privacy concerns
With Cursor, we also have to think about privacy concerns. Cursor.ai has achieved SOC 2 Type II certification and commits to annual security penetration testing by third parties. The platform functions by sending users’ code data to various cloud services to power its AI features, with different handling procedures based on whether Privacy Mode is enabled.
When using Cursor, our code data travels through a network of major cloud service providers and AI model providers, including AWS, Microsoft Azure, Google Cloud Platform, Fireworks, OpenAI, Anthropic, and others. These services may process or temporarily store our code to generate AI responses. However, Cursor has established zero data retention agreements with several AI providers to enhance privacy protection.
If we’re concerned about data privacy, we can enable Privacy Mode, use
.cursorignorefiles to prevent specific files from being indexed, and be aware
that we can completely delete our account and associated data at any time.
While the service necessarily requires sending our code to cloud servers for AI
processing, the combination of zero-retention agreements, data encryption, and
the Privacy Mode guarantee provides reasonable protection for most development
scenarios.
Windsurf (part of Codeium)
Windsurf is another cloud based editor with integrated AI capabilities. It’s part of Codeium, a platform that offers a range of AI-powered development tools for teams and enterprises. Windsurf is very similar to Cursor in terms of capabilities.
How it works
Windsurf models are the following: GPT-4o, Claude 3.5 Sonnet, Claude 3.7 Sonnet , Claude 3.7 Sonnet(Thinking), DeepSeek-V3 , DeepSeek-R1 , o3-mini (medium reasoning), Gemini 2.0 Flash , Gemini 2.5 Pro and Cascade Base.
Cascade base powers the agentic code generation capabilities of Windsurf, allowing it to generate code by just writing continue in their chat. In practice, I’ve seen the same results as using the chat of Cursor. The code generated is very similar and the context awareness is very good.
Cool features
One cool thing about Windsurf is that it offers integrations with MCP (Model Context Protocol) that allow us to use our LLM to access external tools and services. MCP allows Windsurf to draw context from diverse information sources simultaneously. When working on a complex project, the AI can access our local codebase, documentation from the web, private company knowledge bases, and other specialized data sources all through a unified protocol.
Privacy concerns
In terms of privacy, Windsurf inherits Codeium’s privacy approach, which includes a"Zero Data Retention" mode enabled by default for teams and enterprises. This means code or codedata shouldn’t be stored in plaintext on their servers or by their subprocessors, but always remember that our code is still transmitted to Codeium’s servers for processing.
Windsurf relies on various AI model providers like OpenAI, Anthropic, and Google Cloud. While Codeium claims to have zero data retention agreements with these providers, we are still dependent on these third parties honoring their commitments. Also, The Model Context Protocol (MCP) that makes Windsurf powerful also introduces privacy considerations. Since it’s designed to connect to multiple data sources and tools, it potentially increases the pathways through which our data flows.
The fundamental privacy question with Windsurf is the same as with the other tools; how comfortable are we with our proprietary code being processed (even temporarily) by external systems?
More about privacy - where is all the data going? What is the tool doing with it?
GitHub Copilot holds an impressive array of certifications including SOC 1/2/3, ISO 27001:2013, CSA STAR Level 2, and TISAX (specific to the automotive industry). This comprehensive certification profile demonstrates strong security controls across financial reporting, information security management, and industry-specific requirements.
Cursor.ai has achieved SOC 2 Type II certification and conducts annual penetration testing, which covers essential security controls but is less comprehensive than Copilot’s certification suite. Windsurf offers SOC 2 Type II certification, FedRAMP High creditation, and annual penetration testing. While it has fewer total certifications than Copilot, the FedRAMP High accreditation is particularly significant for government and regulated industries.
GitHub Copilot distinguishes between Business/Enterprise data (which they state is not used for training) and individual user data, with less clarity about how individual data is handled by default. Cursor.ai offers “Privacy Mode” which prevents code storage but still allows processing, with this mode automatically enforced for team members. Windsurf features “Zero Data Retention” mode enabled by default for all teams and enterprises, with clear infrastructure separation for privacy-mode requests through parallel processing systems.
GitHub Copilot provides limited information about data residency options in the material reviewed. Cursor.ai mentions servers in US, Asia (Tokyo), and Europe (London)to support global users with lower latency. Codeium servers are specific to regions including US, EU (Germany), and FedRAMP environments, with clear data residency controls to meet regulatory requirements in different jurisdictions.
It’s clear that all three platforms take security seriously but with different approaches. GitHub Copilot has the most comprehensive third-party security validations. The fundamental difference is in data control. GitHub Copilot keeps all processing in their cloud with limited transparency about individual user data. Cursor and Windsurf offers strong privacy controls but still requires sending code to their infrastructure.
Learning concerns
As a programming learner, consider using AI coding tools as assistants rather than replacements for our own thinking. Try solving problems independently first, then compare our solution with AI suggestions to learn alternative approaches. Use AI to explain unfamiliar concepts or to help when you’re truly stuck, but avoid the temptation to generate solutions for every challenge you face. Balance the convenience of AI assistance with deliberate practice that develops our independent coding abilities.
Remember that professional programmers aren’t just code producers; they’re problem solvers who understand the underlying principles and can adapt to new languages, frameworks, and challenges. These fundamental skills come from hands-on practice and working through difficulties, not from outsourcing the thinking process to AI.