The short answer. Source code is the most common sensitive data sent to generative AI apps because developers use AI to debug, explain and rewrite code, and the fastest way to ask is to paste it in. Netskope found source code made up 42% of generative AI data policy violations in 2025 (Netskope, 2026). Bans push the same behavior into personal accounts. What works is a pairing: monitoring and blocking at the network edge for public tools, and a sanctioned AI workspace that answers from your own material under your own access rules.
Source code tops the list
Netskope’s Cloud and Threat Report 2026 broke down generative AI data policy violations by data type. Source code led at 42%, followed by regulated data at 32% and intellectual property at 16%, with passwords and keys making up the rest. In the average organization, 3% of generative AI users committed an average of 223 data policy violations per month, and 50% of organizations lacked enforceable data protection policies for generative AI apps (Netskope, Cloud and Threat Report 2026).
The cost shows up in breach data. IBM found that shadow AI added an average of $670,000 to breach costs, and 63% of organizations had no AI governance policy or were still developing one (IBM and Ponemon Institute, 2025).
Why developers keep pasting code
Developers paste code for a practical reason. A stack trace or a legacy module nobody documented is exactly the problem generative AI is good at, and the public tool is one tab away. The developer is doing their job faster. The code just left the building on the way.
Samsung learned this in 2023. After engineers uploaded sensitive internal source code to ChatGPT, the company restricted staff use of generative AI tools on company devices (Bloomberg, May 2023).
Why bans don't hold
A ban without an alternative moves the behavior somewhere you can’t see it. Netskope’s data shows the direction organizations are taking: the share of AI users on personal AI apps fell from 78% to 47% over the year, as companies moved staff to managed tools (Netskope, 2026). The organizations making progress are replacing, not just prohibiting.
Controls that work
| Control | What it does | Where it lives |
|---|---|---|
| Data loss prevention for AI traffic | Detects and blocks sensitive code and regulated data headed to public AI apps | Network edge, DLP or CASB |
| Managed accounts only | Keeps AI use inside accounts the company controls | Identity and procurement |
| A sanctioned AI workspace | Gives staff a better option for internal questions | Your approved AI platform |
| Access enforced where the AI reads | Stops an internal AI from surfacing material the user isn't cleared for | Your approved AI platform |
| Answers tied to named internal sources | Shows which document each answer came from | Your approved AI platform |
| Protection against extraction attempts | Catches prompts written to pull sensitive content out | Your approved AI platform |
The first control watches the door. The rest make the door less necessary. You need both.
Where Mithra fits
Mithra is the sanctioned-workspace half of that pairing. It doesn’t monitor or block traffic to public AI tools; that’s the job of your DLP or CASB. What it does is give employees an AI they can ask about internal material, with controls a public chatbot doesn’t have.
Documents go into permissioned Library spaces, and space membership decides whose questions can reach which documents, so an architecture document shared with one team isn’t an answer for the whole company. Each file is tokenized on upload with a Token ID, a timestamp, its owner and its metadata. Every answer shows the documents it relied on, and View Evidence opens each one.
AI Threat Protection runs on every query, always on, and detects prompt injection, attempts to extract sensitive data, hidden instructions embedded in uploaded files and output manipulation. Documents are treated as data, never as instructions. Admins can block restricted topics with a set message and exclude documents from retrieval by metadata value, effective immediately across the organization. Each document’s Data Lineage view logs when it was cited in a response or viewed as evidence, so owners can see how their material is being used.
Mithra works on the files you upload: PDF, DOCX, TXT, XLSX, XLS and CSV, up to 150MB each. It isn’t a code assistant and doesn’t connect to your repositories.
Frequently asked questions
Why is source code the most common data leaked to AI tools?
Because debugging and refactoring code are among the most common reasons developers use generative AI, and pasting the code is the fastest way to ask. Netskope measured source code at 42% of generative AI data policy violations (Netskope, 2026).
Does banning ChatGPT stop code leaks?
Not on its own. Without a sanctioned alternative, use moves to personal accounts and unmanaged devices. Pair a ban or restriction with DLP for AI traffic and a managed AI tool staff will actually use.
How do you stop employees from putting IP into public LLMs?
Block sensitive data headed to public AI apps at the network edge and require managed accounts. Then give staff a governed AI workspace for internal questions, with access controls enforced where the AI reads.
About Mithra Technologies Inc.
Mithra Technologies Inc. is building the trust infrastructure for enterprise AI. Centered on its cryptographic Single Source of Truth (SSOT®), Mithra enables organizations to establish verifiable AI through trusted enterprise content, cryptographic proof, complete data lineage, provenance and governance. Designed to work across any AI model, Mithra helps enterprises deploy AI that is more transparent, explainable, auditable and resilient while reducing the risks associated with misinformation, unauthorized data usage, cyber threats and the emerging quantum era.