A new study reveals that encryption methods in AI inference logs can be undermined, risking exposure of personal and authentication information from systems built on Anthropic, OpenAI, and Google APIs, raising urgent questions about enterprise AI data security.
A new study has raised fresh concerns about the way major AI providers handle inference logs, suggesting that data routed through applications built on Anthropic, OpenAI and Google APIs could expose personal and authentication information if those logs are stored or shared carelessly. The research, posted on Hugging Face, says the three companies’ frontier models all use a form of “encrypted inference” intended to hide the internal reasoning process before a final answer is returned to a client. But the study found that this protection can be undermined because the encrypted blocks are interoperable across models, creating a path for one system to decode another’s output.
In the experiment, researchers showed that a stronger Anthropic model could produce an encrypted block that a weaker Claude model could then decrypt. The same structural weakness was also observed in OpenAI’s GPT family and Google’s Gemini line, according to the study summary. By examining 6,708 publicly available agent operation records, the researchers say they decrypted 315,320 inference blocks and identified 367 instances of personally identifiable information, as well as 182 items of authentication data, including API keys and passwords.
The findings matter well beyond laboratory conditions because many companies now use enterprise AI systems to automate customer support and internal work. Amazon Web Services describes enterprise AI as the large-scale adoption of AI in organisations, with governance, compliance and data handling among the main concerns. In South Korea, firms including Samsung Electronics, LG Electronics, LG CNS, Samsung SDS, Krafton and Toss have all been reported to use ChatGPT-based APIs in business processes, while banks such as Shinhan Bank are moving call-centre operations towards AI contact centre systems.
The risk is especially acute for multi-turn chatbots and agentic services, which must preserve context across a session and may therefore pass encrypted inference blocks back and forth repeatedly. That design makes them more exposed if session logs are retained, copied or shared without strict controls. The study has added urgency to questions over log storage, access management and whether regulators should set clearer rules for AI systems that handle sensitive business and customer data.
South Korean industry has also been pushing ahead with sovereign AI projects and new enterprise tools, including local model and hardware collaborations intended to strengthen domestic AI capability. That broader push now sits alongside a sharper warning: as AI becomes more deeply embedded in corporate workflows, the security of the logs it generates may be as important as the model itself.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
