OpenAI Deploys ChatGPT‑Powered Surveillance Tool to Track and Stop Leakers in Real Time

OpenAI’s internal security team rolled out the system in late 2024, feeding a custom‑tuned ChatGPT instance with every public article that mentions the company while granting the model read‑only access to Slack archives, email logs and document repositories, according to The Information. The AI then parses each leak‑related phrase—project codenames like “Q*” or “Project Strawberry,” specific financial figures, and even distinctive wording—against the trove of internal communications. Within minutes it can flag which files contain the exposed data, which employees had legitimate access, and which users have used similar language in private chats, compressing what used to be a weeks‑long manual investigation into a short list of suspects.

The move reflects a broader escalation in OpenAI’s response to a “metronomic” stream of disclosures over the past 18 months, ranging from product roadmaps to executive departures, as chronicled by the Moth post on March 1. Earlier attempts to curb leaks included aggressive NDAs that clawed back vested equity from departing staff who spoke publicly. After a backlash in mid‑2024, OpenAI announced it would stop enforcing those provisions, but the new AI‑driven surveillance tool signals a shift from contractual deterrents to technological enforcement. Employment lawyers and the Electronic Frontier Foundation have warned that such deep‑linguistic analysis—matching authorial fingerprints across thousands of messages—ventures into legal gray zones that privacy statutes have yet to address, a concern echoed in the same report.

The practical impact of the system remains opaque. OpenAI has not confirmed whether any employee has been identified or disciplined as a result, and the company declined to comment on the tool’s efficacy when approached for comment. However, the mere existence of an AI that can cross‑reference every external news story with internal chatter creates a chilling effect, as the Moth analysis notes: “If you know an AI is reading your Slack messages and comparing them to every tech news article published that week, you think twice before messaging a reporter.” That hesitation could suppress legitimate whistleblowing, even though federal and state protections shield employees who report illegal conduct or safety violations—especially critical in a field where misaligned AI systems are framed by OpenAI itself as existential risks.

The surveillance apparatus also underscores a paradox at the heart of OpenAI’s business model. The company markets ChatGPT as a productivity enhancer for the global workforce while simultaneously weaponizing the same technology to police its own staff. This duality mirrors broader industry trends; California employers can legally scan corporate communications with adequate notice, yet the deployment of sophisticated generative AI for author attribution pushes the boundary far beyond traditional keyword filters. The Verge recently highlighted how ChatGPT was tricked into extracting sensitive Gmail data, illustrating the broader vulnerability of AI tools when repurposed for surveillance.

OpenAI’s recent talent exodus adds another layer of context. Senior safety researchers—including Ilya Sutskever, Jan Leike and Daniel Kokotajlo—have departed, some forfeiting equity worth over a million dollars rather than sign non‑disparagement agreements, according to the same source. Their exits were not driven by petty grievances but by concerns over the company’s direction and safety protocols. The new surveillance system, therefore, arrives at a moment when internal dissent is already high, raising questions about whether the tool will be used to protect proprietary information or to silence dissenting voices. As the AI industry watches, OpenAI’s gamble on AI‑powered internal policing may set a precedent that reverberates across Silicon Valley, reshaping the balance between corporate security and employee rights.