Claude Used to Breach Rival OpenAI... Internal Systems Infiltrated Within 72 Hours

[Financial News] A case has emerged in which Anthropic’s artificial intelligence (AI) model, Claude, was used to infiltrate rival OpenAI’s internal systems. By leveraging the latest AI model, the attackers took less than 72 hours to progress from discovering a vulnerability to infiltrating the internal system.
According to Hacktron, an AI security research firm, on the 18th local time, researchers successfully used Claude Opus 5 to take control of the ChatGPT and Codex accounts used by OpenAI employees.
The researchers first accessed a server through a vulnerability in OpenAI’s help forum and obtained the employees’ account information. They then used those credentials to access OpenAI’s internal system, known as Monorepo, and left evidence that they had successfully infiltrated it. The system is reportedly believed to store confidential information related to core algorithms.
Improvements in AI model performance significantly shortened the time required for the attack. The researchers initially tried to use Opus 4.8 to write attack code but failed to make progress. However, after using the newly released Opus 5, they succeeded in creating an attack script in just a few hours.
Less than 72 hours elapsed between the initial discovery of the vulnerability and the infiltration of the internal system. Hacktron explained that AI models are rapidly becoming more powerful, saying, "Although the agent’s work took several days, the human researchers spent only a few hours on it."
The researchers reported the infiltration to OpenAI, which confirmed the incident and paid a $6,500 reward, approximately 9 million won. OpenAI also completed security measures for the affected systems within 14 hours of receiving the report.
As the possibility of AI-assisted hacking grows, the industry is also strengthening its security responses. OpenAI recently assigned 25% of its entire engineering staff to system defense at the direction of President Greg Brockman.
The cybersecurity industry is focusing on the fact that AI can significantly reduce the human involvement and time required to search for vulnerabilities and write attack code. Anthropic CEO Dario Amodei has also raised concerns about the risks of AI technology and called for the industry as a whole to slow the pace of development.
[email protected] Yoon Hong-jip Reporter