Researchers Accessed OpenAI Systems via Claude AI in 72 Hours, Report Says

Researchers accessed OpenAI systems with the help of Anthropic’s Claude AI within 72 hours, according to a report from Chosun Ilbo. The headline carries three elements: the researchers, the Claude model, and the 72-hour window.

The time frame is the most informative part of the report. Security testing against large online systems has traditionally been heavy work requiring specialist skills and weeks of effort. A three-day window, if accurate, points to how far frontier models can compress that process.

Public information does not yet make clear whether the access happened inside an authorized red-team engagement or a bug bounty program, or outside any such framework. That distinction matters: one is routine collaboration between vendors and the security community, the other raises immediate legal and ethical questions.

Handing systems to outside researchers for security testing has become routine practice for model vendors, and bringing AI tools into that loop is increasingly common. What makes this item notable is that it places two of the most closely watched names in AI — Claude and OpenAI — inside the same security scenario.

For users and enterprise customers, the takeaway is the duality of the same capability: one model can be used to harden a system or to look for its gaps. A 72-hour timeline suggests the defender’s response window is shrinking, and that security teams should assume adversaries have access to the same class of tools.

Open questions remain: what exactly was targeted, whether the work was authorized, whether OpenAI has responded, and whether AI-assisted security testing of this kind becomes standard practice.