▲ Dario Amodei, CEO of Anthropic
A case has been discovered where competitor OpenAI was hacked using Anthropic's artificial intelligence (AI) model, Claude.
AI security research firm Hacktron announced on the 18th (local time) that its researchers successfully took control of OpenAI employee accounts for the ChatGPT AI chatbot and the Codex coding tool by using the "Claude Opus 5" model.
They first succeeded in accessing the server through a vulnerability in OpenAI's help forum, where they secured account information for OpenAI employees.
They then used this account information to access OpenAI's internal system, known as a "monorepo," and left evidence of their infiltration.
This system is known to be a repository where core algorithmic secrets are stored.
They initially attempted to write attack code using "Opus 4.8" without success, but stated that when they used "Opus 5," which Anthropic had just released at the time, they were able to successfully create the attack script in just a few hours.
It took less than 72 hours from the initial discovery of the vulnerability to infiltrating the internal system.
Regarding the time required for this hack, Hacktron explained, "While the agent's work took a few days, the time spent by human researchers was merely a few hours," adding, "As new AI models emerge, their performance is becoming increasingly powerful."
They reported their infiltration to OpenAI, and after verifying the report, OpenAI paid a bounty of $6,500 (approximately 9 million won).
OpenAI also completed system security measures within 14 hours of receiving the report.
OpenAI recently reassigned 25% of all engineers to system defense tasks under the instruction of President Greg Brockman.
The cybersecurity industry is facing a period of upheaval as AI-driven hacking cases increase.
Some of OpenAI's models previously caused shock by escaping isolated environments and launching cyberattacks against external targets even without explicit human instructions.
Anthropic CEO Dario Amodei has argued for adjusting development speeds across the entire industry, sparking ongoing debate within the AI sector.
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.