Research has revealed that personal data could be exposed through inference logs from companies that have built AI services for call centers and internal automation using APIs from Anthropic, OpenAI, and Google.
A paper titled "Stealing Inference Traces from Proprietary LLM APIs," published on Hugging Face on August 12, identified structural flaws in the 'encrypted inference' blocks returned to clients by the frontier models of the three companies.
Encrypted inference is a method where the AI model transmits its internal thought process in an encrypted form to the client, preventing competitors from observing this process to train their own models and ensuring that the reasoning process is not exposed to users. All three companies adopted this approach for security reasons.
The issue arises from the fact that these encrypted blocks are interoperable across the same AI models. When a higher model requests to decrypt an encrypted block from a lower model, it is returned in plaintext. The paper demonstrated this flaw through an experiment where an encrypted block generated by Anthropic's Claude Opus 4.8 was decrypted by the relatively weaker Claude Haiku 4.5. Similar structural flaws were confirmed in OpenAI's GPT series and Google's Gemini series.
The research team decrypted 315,320 inference block encryptions from 6,708 actual agent operation records available in public repositories. Among these, they identified 367 instances of personal identifiable information (PII) and 182 instances of authentication information, such as API keys and passwords. Given that this data was extracted from publicly available logs, the implications are significant.
Domestic companies utilizing the APIs from the three firms are also at risk of personal data exposure. Many South Korean AI service companies are integrating the APIs from Anthropic, OpenAI, and Google for call center consultations and internal automation.
Samsung Electronics has implemented ChatGPT-based APIs across software development, product development, and marketing. Other companies, including LG Electronics, LG CNS, Samsung SDS, Krafton, and Toss, are also using ChatGPT-based APIs for internal automation. The financial sector, including Shinhan Bank, is increasingly automating call center consultations through AI contact centers (AICC).
Particularly vulnerable are multi-turn chatbots and agent-type services that require maintaining conversational context, as they inherently exchange encrypted inference blocks in each session. This vulnerability necessitates a review of session log storage and sharing practices, along with the development of response measures by relevant authorities.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.