This parallels psychologist Zick Rubin's "stranger-on-a-train" phenomenon: people overshare with strangers they'll never meet again. For corporate information security, the LLM is the train. This article describes the four core privacy challenges Guardora is addressing.
Corporate information security is increasingly concerned about the growing number of cases where employees share sensitive data with ChatGPT, Gemini, Copilot, and other less-known large language models.
Lawyers create summaries of contracts containing counterparty data, technical specifications, prices, and special terms. Programmers edit code critical to system architectures. Translators work with texts full of sensitive information.
On the other hand, companies are trying to integrate AI assistants and replace traditional machine learning models with LLMs.
Here are some use cases we've encountered at Guardora:
- Listening to sales team recordings to score and analyze their interactions with clients.
- Understanding the tone of client interactions in real-time. For example, if a client becomes more irritable, the system can switch them to a higher-ranking specialist who can address their concerns and defuse negative sentiments.
- Clustering users based on interests and other criteria.
- Creating document search and analysis systems with tasks like summarization, keyword exclusion, and personal data protection.
- Generating text, images, music, and other media using sources with unique proprietary content.
- Localizing videos from the original language to others.
- Developing new product clusters—assistants or copilots—that support development teams, enhancing productivity and reducing time-to-market. They minimize errors and handle tasks that are often disliked (e.g., tests, documentation, training juniors). These tools primarily focus on generating, verifying, and debugging code across various tests.
However, the overarching challenge in this vast market is using LLMs while ensuring the data remains accessible only to the data owner. How can we protect data from third parties, cloud service providers, or malicious actors throughout its transmission, storage, model training, quality verification, and result retrieval?
This challenge is precisely what Guardora aims to address. Our solutions already ensure data security for some use cases throughout its entire journey, including ML model training, quality checks, inference, and sometimes even safeguarding the model itself as an intellectual property.
At Guardora, we focus on ensuring the confidentiality of requests to ML models and we want to do the same with LLMs and generative models.
Here is a list of current challenges we need to solve before launching our first prototypes:
- Safeguarding the confidentiality of training datasets during the initial training of LLMs and generative models.
- Preventing unauthorized third-party usage of LLMs and generative models trained from scratch.
- Protecting the confidentiality of data used to fine-tune pre-trained LLMs and generative models.
- Securing the confidentiality of queries submitted to LLMs and generative models.
If this topic interests you as a user or developer, join our Discord community and participate in the discussion of these pressing issues.
Frequently Asked Questions
What is the "stranger-on-a-train" phenomenon in the context of LLMs?
The "stranger-on-a-train" phenomenon was originally coined by social psychologist Zick Rubin to describe the tendency of people to share personal information with strangers they're unlikely to meet again — fellow passengers on a long train ride, for example. Applied to Large Language Models: employees often share sensitive corporate data (contracts, source code, client details) with ChatGPT, Gemini, or Copilot in ways they would never share with a human stranger, because the LLM feels like a one-time conversation even when, in practice, the data leaves the company perimeter and may be used for model training, logged for compliance, or processed by third-party providers.
What kinds of sensitive data do employees commonly share with ChatGPT, Gemini, and Copilot?
Based on commonly reported scenarios: lawyers submit contract summaries containing counterparty data, prices, and special terms. Programmers paste source code from architecture-critical systems for debugging or refactoring suggestions. Translators process texts containing trade secrets, M&A details, or personal information. Sales and customer success teams transcribe and analyze client call recordings. Product teams feed proprietary documents for summarization and analysis. Analysts upload financial data for interpretation. None of these operations should leave the organization's perimeter under standard corporate information security policies.
Can my corporate data be exposed when I use ChatGPT for work tasks?
Depends on which deployment of the LLM and what plan. Consumer ChatGPT (free tier) historically logged conversations and used them for model improvement — though OpenAI now offers opt-outs and Enterprise tiers that don't train on customer data. Microsoft Copilot, when integrated through Microsoft 365 Enterprise plans, has data-protection guarantees. However: the model provider sees the prompts in cleartext on its infrastructure. Truly confidential workloads require either on-premise LLM deployment, or cryptographic protection of the inference request itself — the latter being the privacy-preserving ML approach Guardora researches.
What are the main privacy challenges in deploying LLMs commercially?
Guardora identifies four core challenges.
- Training data confidentiality — protecting the proprietary datasets used to train an LLM or generative model from initialization.
- Unauthorized model use — preventing third parties from running inference on a model that was trained from scratch on someone else's data.
- Fine-tuning data confidentiality — protecting the sensitive data used to fine-tune a pre-trained foundation model (e.g., domain-specific corporate knowledge).
- Query confidentiality — protecting the prompts and outputs of inference requests so the model provider, cloud, or eavesdroppers cannot see them.
How can companies use LLMs without leaking confidential data?
Three primary approaches.
- On-premise LLM deployment — running open-source models (Llama, Mistral, Qwen variants) inside the company's own infrastructure. Most expensive but most controllable.
- Enterprise tiers with contractual guarantees — ChatGPT Enterprise, Copilot for M365, Gemini Enterprise — model provider commits not to train on customer data, but still sees prompts in cleartext.
- Privacy-preserving LLM inference — cryptographic techniques (homomorphic encryption, secure multi-party computation, trusted execution environments) so the provider cannot see prompts in cleartext.
This third path is still maturing and is where Guardora is researching.
What is confidential LLM inference?
Confidential LLM inference is the cryptographic protection of the prompt-and-response pipeline so that the entity running the LLM (the model provider, cloud, or any third party) cannot see the user's actual prompt or the model's actual response in cleartext. Several techniques are in active research:
- homomorphic encryption of the input/output (very slow but mathematically tight),
- trusted execution environments (TEEs) like Intel TDX, AMD SEV-SNP, NVIDIA Confidential Compute (hardware isolation, faster but trust the hardware), and
- secure multi-party computation (splits computation across parties).
Production-grade confidential LLM inference at GPT-4 scale is still nascent in 2026.
Who is Zick Rubin and what is the original "stranger-on-a-train" theory?
Zick Rubin (born 1944) is an American social psychologist best known for his research on attraction and self-disclosure. He developed the "Rubin Scale" measuring liking versus loving, and his work on self-disclosure described the so-called stranger-on-a-train phenomenon: under conditions of anonymity and one-time interaction, people often disclose more personal information than they would to people they expect to see again. The Guardora article applies this human-behavior insight to the LLM era — the LLM provides the illusion of one-time anonymous interaction, inviting oversharing.
What is Guardora doing about LLM privacy?
Guardora's core production products today are Guardora VFL (Vertical Federated Learning for tabular ML — banks, insurers, healthcare) and Guardora FFT (federated fine-tuning, including for foundation models). LLM-specific privacy is an active research area — Guardora has publicly identified four challenges to solve before launching commercial LLM-privacy products: training confidentiality, anti-unauthorized-use, fine-tuning data protection, and query confidentiality. For active discussion of the LLM privacy roadmap, join the Guardora Discord community linked from the article. Production-ready confidential LLM products are on Guardora's roadmap rather than commercially available today.
