OpenAI Uses ChatGPT to Identify Employee Leakers [Model Behavior]

[00:00] Chad Thompson: Welcome to Model Behavior. [00:02] Chad Thompson: Model Behavior examines how AI systems are built, deployed, and operated in real professional environments. [00:09] Chad Thompson: Joining us today is Chad Thompson, who provides a systems-level perspective on AI, automation, and security. [00:17] Chad Thompson: Chad, it is good to have you. [00:18] Nina Park: Thanks, Nina. [00:19] Nina Park: We are starting with a report from yesterday regarding OpenAI. [00:23] Nina Park: The company is reportedly using a custom internal version of ChatGPT to identify employees who leak confidential information. [00:31] Nina Park: According to the information, security personnel run published news stories through the model, [00:37] Nina Park: which has access to internal slat logs, emails, and documents to cross-reference specific leaked details. [00:44] Thatcher Collins: It is a clear example of using LOMS for internal telemetry analysis. [00:49] Thatcher Collins: I mean, from a systems perspective, the ability to automate the matching of unstructured communication data against external reports is a powerful security tool. [01:00] Thatcher Collins: However, it shifts the role of the AI from a productivity assistant to a surveillance mechanism within the corporate environment. [01:08] Chad Thompson: Thatcher, on the consumer side, we've seen a very different use of these auditing capabilities. [01:14] Chad Thompson: A report yesterday highlighted a patient, Matt Rosenberg, who used Anthropics Claude [01:19] Chad Thompson: to audit a $195,000 hospital bill. [01:23] Chad Thompson: Claude identified that the hospital had unbundled procedures that Medicare requires to be [01:29] Chad Thompson: billed as a single package, eventually helping negotiate the bill down to $32,000. [01:36] Thatcher Collins: Mm-hmm. [01:36] Thatcher Collins: Nina, it's a massive win for patient advocacy. [01:40] Thatcher Collins: It is also worth noting that Anthropic made these specific advanced tools, including file [01:47] Thatcher Collins: creation, connectors, and custom skills, available to all free users earlier this week. [01:53] Thatcher Collins: they are clearly positioning Claude as a tool for high-utility document analysis [01:59] Thatcher Collins: and specialized tasks without the need for a subscription. [02:03] Thatcher Collins: The medical billing case is significant because it required Claude to act as a specialized auditor, [02:08] Thatcher Collins: comparing CPT codes against federal regulations. [02:12] Thatcher Collins: This move towards skills and tools like Claude Code [02:15] Thatcher Collins: suggests we are transitioning from simple chatbots to autonomous agents [02:20] Thatcher Collins: that can navigate complex bureaucratic systems. [02:23] Chad Thompson: Turning to industry shifts, earlier this week, Elon Musk held an all-hands meeting for XAI. [02:31] Chad Thompson: The company has now split into four specialized teams, Grok Main, Coding, Imagine, and a new simulation project called Macrohard. [02:41] Chad Thompson: Musk also discussed long-term plans for orbital data centers and lunar satellite factories to explore deep space. [02:50] Nina Park: Musk's vision is certainly expansive, but we are also seeing immediate practical enterprise deployments. [02:59] Nina Park: For instance, the developer Evron recently integrated QN and Yandex GPT into their internal systems to automate HR resume parsing. [03:11] Nina Park: They reported a 90% reduction in manual salary lookups by using LLMs to normalize unstructured data. [03:20] Thatcher Collins: Both MacroHard and the Everone case show that the next phase of AI is about deep integration into business infrastructure. [03:28] Thatcher Collins: Whether it is simulating entire software firms or just cleaning up HR data, the goal is reducing manual friction in the system. [03:37] Thatcher Collins: The focus is shifting from what the AI can say to what the AI can do within a professional workflow. [03:44] Chad Thompson: Thank you for listening to Model Behavior, a neural newscast editorial segment, [03:50] Chad Thompson: mb.neuralnewscast.com. [03:55] Chad Thompson: Neural Newscast is AI-assisted, human-reviewed. [03:59] Chad Thompson: View our AI transparency policy at neuralnewscast.com.